Expecting the Unexpected:Training Detectors for Unusual Pedestrians with Adversarial Imposters原文
Expecting the Unexpected:Training Detectors for Unusual Pedestrians with Adversarial Imposters
Shiyu Huang
Tsinghua University
Beijing, China
[email protected]
Deva Ramanan
Carnegie Mellon University
Pittsburgh, USA
[email protected]
Abstract
As autonomous vehicles become an every-day reality,
high-accuracy pedestrian detection is of paramount practical importance. Pedestrian detection is a highly researched
topic with mature methods, but most datasets focus on common scenes of people engaged in typical walking poses on
sidewalks. But performance is most crucial for dangerous
scenarios, such as children playing in the street or people
using bicycles/skateboards in unexpected ways. Such “inthe-tail” data is notoriously hard to observe, making both
training and testing difficult. To analyze this problem, we
have collected a novel annotated dataset of dangerous scenarios called the Precarious Pedestrian dataset.
Even given a dedicated collection effort, it is relatively
small by contemporary standards ( ≈ 1000 images). To allow for large-scale data-driven learning, we explore the use
of synthetic data generated by a game engine. A significant
challenge is selected the right “priors” or parameters for
synthesis: we would like realistic data with poses and object configurations that mimic true Precarious Pedestrians.
Inspired by Generative Adversarial Networks (GANs), we
generate a massive amount of synthetic data and train a discriminative classifier to select a realistic subset, which we
deem the Adversarial Imposters. We demonstrate that this
simple pipeline allows one to synthesize realistic training
data by making use of rendering/animation engines within
a GAN framework. Interestingly, we also demonstrate that
such data can be used to rank algorithms, suggesting that
Adversarial Imposters can also be used for “in-the-tail”
validation at test-time, a notoriously difficult challenge for
real-world deployment.
(a) (b) (c)
Figure 1: (a) Examples from our novel Precarious Pedestrian Dataset of dangerous, but rare pedestrian scenes. One
important scenario is that of pedestrians on their phone (row
2, col 2), who may not be adequately aware of their surroundings. (b) Examples in Caltech Dataset tend not to
capture such rare scenarios. (c) Examples from a set of
Adveserial Imposters, which are synthetic images that are
adversarially-trained to mimic the set of Precarious Pedstrians. We demonstrate that such images can be used to both
train and evaluate robust pedestrian recognition systems targeting such dangerous scenarios.
1. Introduction
There’s no software designer in the world that’s ever going to be
smart enough to anticipate all the potential circumstances an
autonomous car is going to encounter. The dog that runs out into
the street, the person who runs up the street, the bicyclist, the
policeman or the construction worker.
C. Hart, Chairman of National Transport. Safety Board
As autonomous vehicles become an every-day reality,
high-accuracy pedestrian detection is of paramount practical importance. Pedestrian detection is a highly researched
arXiv:1703.06283v2 [cs.CV] 10 Apr 2017
topic with mature methods, but most datasets focus on
“everyday” scenes of people engaged in typical walking
poses on sidewalks [9, 6, 13, 12, 48]. However, perhaps
the most important operating point for a deployable system is its behaviour in dangerous, unexpected scenarios,
such as children playing in the street or people using bicycles/skateboards in unexpected ways.
Precarious Pedestrian Dataset: Such “in-the-tail” data is
notoriously hard to observe, making both training and evaluation of existing systems difficult. To analyze this problem, we have collected a novel annotated dataset of dangerous scenarios called the Precarious Pedestrian Dataset.
Even given a dedicated collection effort, it is relatively
small by contemporary standards ( ≈ 1000 images). To explore large-scale data-driven learning, we explore the use of
synthetic data generated by a game engine. Synthetic training data is an actively explored topic because it provides a
potentially infinite well of annotated data for training datahungry architectures [24, 30, 14, 19, 40, 42]. Particularly
attractive are approaches that combine a large amount of
synthetic training data with a small amount of real data (that
may have been difficult to acquire and/or label).
Challenges in Synthesis: We see two primary difficulties
with the use of synthetic training data. The first is that not
all data is created “equal”: when combining synthetic data
with real data, synthesizing common scenes may not be particularly useful since they will likely already appear in the
training set. Hence we argue that the real power of synthetic
data is generating examples “in-the-tail”, which would otherwise have been hard to collect. The second difficulty
arises in building good generative models of images, a notoriously difficult problem. Rather than building generative
pixel-level models, we make use of state-of-the-art rendering/animation engines that contain an immense amount of
knowledge (about physics, light transfer, etc.). The challenge of generative synthesis then lies in constructing the
right “priors”, or scene-parameters, to render/animate. In
our case, these correspond to body poses and spatial configurations of people and other objects in the scene.
Adversarial Imposters: We address both concerns
with a novel variant of Generative Adversarial Networks
(GANs) [17], a method for synthesizing data from latent
noise vectors. Traditional GANs learn generative feedforward models that process latent noise vectors, typically
from a fixed known prior distribution. Instead, we fix the
feedforward model to be a rendering engine, but use an adverserial framework to learn the latent priors. To do so,
we define a rendering pipeline that takes an input a vector of scene parameters capturing object attributes and spatial layout. We use rejection sampling to construct a set
of scene parameters (and their associated rendered images)
that maximally confuse the discriminator. We call such examples Adversarial Imposters, and use them within a sim-
(a) (b)
Figure 2: (a) Scene that’s built for generating synthetic images. (b) 3D models that we use in this project.
ple pipeline for adapting detectors from synthetic data to the
world of real images.
RPN+: We use our dataset of real and imposter images to
train a suite of contemporary detectors. We find surprisingly good results with a (to our knowledge) novel variant
of region proposal network (RPN) [49] tuned for particular objects (precarious people) rather than a general class of
objectness detections. Instead of classifying a sparse set of
proposed windows (as nearly all contemporary object detection systems based on RCNN do [38]), this network returns a dense heatmap of pedestrian detections, along with
regressed bounding box location for each pixel location in
the heatmap. We call this detector RPN+. Our experiments
show that our RPN+, trained on real+imposter data, outperforms other detectors trained only on real data.
Validation: Interestingly, we also demonstrate that our Adverserial Imposter Dataset can be used to rank algorithms,
suggesting that our pipeline can also be used for “in-thetail” validation at test-time, a notoriously difficult challenge
for real-world deployment.
Contributions: The contribution of our work is as follows:
(1) a novel dataset of pedestrians in dangerous situations
(Precarious Pedestrians) (2) a general architecture for creating realistic synthetic data “in-the-tail”, for which limited real data can be collected and (3) demonstration of our
overall pipeline for the task of pedestrian detection using a
novel detector. Our datasets and code can be found here:
https://github.com/huangshiyu13/RPNplus.
2. Related work
Synthetic Data: Synthetic datasets have been used to train
and evaluate the performance of computer vision algorithms. Some forms of ground truth are hard to obtain from
hand-labelling, such as optical flow, but easy to synthesize
via simulation [14]. Adam et al . [24] used a 3D game
engine to generate synthetic data and learned an intuitive
physics model to predict falling towers of blocks. Mayer et
al . [30] released a benchmark suite of various tasks using
synthetic data, including disparity and optical flow. Richter
et al . [40] used synthetic data to improve image segmentation performance, but notably do not control the scene as
to explore targeted arrangements of objects. German Ros
et al . [42] used Unity Development Platform to generate a
synthetic urban scene dataset.
3D Models for Detection: A notable application of 3D
computer graphics model in vision has been the modeling of the human body shapes [18, 4, 1, 29, 36, 3, 41].
Moreover, 3D simulation can also been used for car detection [34, 32, 20] and scene understanding [45, 23]. Marin et
al . [29] used a game engine to generate synthetic training
data. Pishchulin et al . [35] used 8 HD cameras to scan human body and built real 3D human models. Then they used
synthetic data and some labelled real data to train pedestrian
detectors. Hattori et al . [19] used 3D modelling software
to build a special scene and randomly put 3D models on a
special background for pedestrian detection. Most of these
works use synthetic data “as-is”, while we analyze statistical differences between synthetic and real data, describing a
pipeline for reconciling such differences through adversarial domain adaption.
Domain Adaptation: Domain Adaptation is a standard
strategy to deal with data across different domains, such as
synthetic versus real. Large synthetic datasets can be used
to bootstrap detectors and then adapted to real data by moving to the target domain distribution. Sun and Saenko [46]
used 3D models to train detectors for real objects. Such
work typically used shallow detectors defined on fixed feature sets, while we focus on gradient-based adaption of
“deep” detection networks (such as RCNN). From this perspective, our work is inspired by approaches for deep domain adaptation [15, 16, 26, 27]. Such work typically assumes that one has access to large amounts of unlabeled
data from the target domain. In our case, assembling a large
target dataset of unlabelled examples (of real Precarious
Pedestrians) is itself challenging, necessitating the need for
alternative approaches that make stronger use of the source
dataset.
Generative Adversarial Nets: GANs [17] are deep networks that can generate synthetic images from latent noise
vectors. They do so by adversarially-training a neural network to discriminate between real versus synthetic images.
Recent works have shown impressive performance in generation of synthetic images [31, 7, 37, 44, 5]. However,
it appears challenging to synthesize high-resolution images
with semantically-valid content. We circumvent these limitations with a rendering-based adversarial approach to image synthesis.
(a) Precarious Dataset (b) Caltech Dataset
Table 1: Constraints of parameters for synthesizing images.
The index and time(normalized) of animations will jointly
decide the gestures of 3D models .
Figure 2 shows the commercial 3D human models
that we use for data generation, consisting of 20 models
spanning different women, men, cyclists and skateboarder
avatars. Because these are designed for game engine
play, each 3D model is associated with characteristic
animations such as jumping, talking, running, cheering and
applauding. We animate these models in a 3D scene with
a 2D billboard to capture the scene background [11], as
shown in Figure 2. Billboards are randomly sampled from
the 1726 background images from INRIA dataset [6] and
a custom set of outdoor scenes downloaded from Internet.
Our approach can generate a diverse set of background
scenes, unlike approaches that are limited to a single virtual
urban city [29].
Scene parameters: To build a large library of synthetic images that will potentially be used for training and evaluation,
we first define a set of parameters and parameter ranges. We
index the set of background images, the set of 3D models,
and the animation frame number for each model. In brief,
the scene parameters include directional light intensity and
direction (capturing sunlight), the background image index,
the number of 3D models, and for each model, an index
specifying the avatar ID and animation frame, as well as a
root position and orientation (rotation in the ground plane).
We assume a fixed camera viewpoint. Note that the root position affects both the location and scale of the 3D model
in the rendered image. All these parameters can be summarized as a variable-length vector z 2 Z , where each vector
corresponds to a particular scene instantiation.
Synthesis: Our generator G ( z ) , or rendering engine, synthesizes an image corresponding to z . Importantly, we can
also synthesize labels L ( z ) for each rendered image, spec-
(a) (b)
Figure 4: (a) Imposter images that are chosen by selector.
(b) Synthetic images that are not in Imposter Dataset.
ifying object type, 3D location, pixel segmentation masks,
etc . In practice, we make use of only 2D object bounding
boxes. Table 1 shows the viable ranges of each parameter.
In addition, we found the following heuristic to simulate
reasonable object layouts: we enforce the maximum overlap between any two 3D models to be 20% (to avoid congestion) and the projected location of the 3D models should
lie within the camera’s field-of-view. These conditions are
straightforward to verify for a given vector z without rendering any pixels, and so can be efficiently enforced though
rejection sampling (i.e., generate a random vector and only
render those that pass these conditions). Unlike Hironori et
al . [19], who generate training data by manually tuning z
to match specific scenes, our approach is not scene specific
and does not require any manual intervention.
Pre-processing: Synthesized images and Precarious Pedestrian images may be of different sizes. We isotropically
scale each image to a resolution of 960 × 720, zero-padding
as necessary. Our experiments also make use of the Caltech Pedestrian benchmark, to which we apply the same
pre-processing.
4. Proposed Method
Domain adaption: In this section, we introduce a novel
framework for adversarially adapting detectors from synthetic training data to real training data. We use x 2 X to
denote an image and y 2 Y to denote its label vector (a set
of bounding box labels). Let p s ( x ; y ) to refer to the distribution of image-label pairs from the source domain (of
synthetic images), and p t ( x ; y ) to refer to the target domain
(of real Precarious Pedestrians). In our problem, we expect large amounts of source samples, but a limited amount
of target ones. We factorize the joint into a marginal over
image appearance and conditional on label given the appearance - e.g., p s ( x ) p s ( y j x ) . Importantly, we discriminatively train a feedforward function f s ( x ) = p s ( y j x ) to
match the conditional distribution. Our central question is
how to transfer feedforward predictors trained from source
samples f s ( x ) to the target domain f t ( x ) .
Fine-tuning: The most natural approach to domain adaption may simply be to fine-tune a predictor f s ( x ) , originally trained on the source, with samples from the target
p t ( x ; y ) . Indeed, virtually all contemporary methods for visual recognition makes use of fine-tuned models that were
pre-trained on Imagenet [43]. We compare to such a strategy in our experiments, but find that fine-tuning works best
when source and target distributions are similar. As we
argue, while rendering engines can produce photorealistic
scenes, it is difficult to specify a prior over scene parameters
that mimic real (Precarious) scenes. We describe a solution
that adversarially learns a prior.
Generators: As introduced in Sec. 3.2, let z 2 Z be a
vector of scene parameters, G ( z ) 2 X be a feedforward
generator function that renders a synthetic image given the
scene parameters, and L ( z ) 2 Y be a function that generates labels from the scene parameters. We can then reparameterize the distribution over synthetic images as a distribution over scene parameters p z ( z ) . We now describe a
procedure for learning a prior p z ( z ) that allows for easier
transfer. Specifically, we learn a prior that fools an adversary that is trying to distinguish samples from the source
and target.
Adversarial generators: To describe our approach, we first
recall a traditional generative adverserial network (GAN):
min
G
max
D
V ( D; G ) = [Gen. Adversarial Net] (1)
E
x s p t ( x ) [log D ( x )] + E z s p z ( z ) [log(1 - D ( G ( z )))]
where the minmax optimization jointly tries to estimate a
discriminator D that can distinguish real versus synthesized
data examples, and the generator G tries to synthesize realistic examples that fool the discriminator. Typically, the
discriminator D ( x ) is trained to output the probability that
x is real (e.g., a real Precarious Pedestrian), while p z ( z )
is fixed to be a zero-mean, unit-variance Gaussian. This
optimization can be performed with stochastic gradient updates, that converge (in the limit) to a fixed point of the minimax problem. We refer the reader to the excellent introduction in [17]. Importantly, the generator must encode complex constraints about the manifold of natural images, that
capture amongst other knowledge the physical properties of
light transport and material appearance.
Adversarial priors: We note that rendering engines can
be viewed as generators that already contain much of this
knowledge, and so we fix G to be a production-quality rendering platform (Unity 3D). Instead, we learn the prior over
parameter vectors in a adversarial manner:
min
I
max
D
V ( D; I ) = [Adversarial Priors] (2)
E
x s p t ( x ) [log D ( x )] + E z s p I ( z ) [log(1 - D ( G ( z )))]
If the generator G is differentiable with respect to z , it is
possible to use backprop to compute gradient updates for
simple prior distributions p I ( z ) , such as Gaussians [22, 39].
This implies that the above formulation of adversarial priors
is amenable to gradient-based learning.
Imposter search: We see two difficulties with directly applying (2) to our problem: (1) It seems unlikely that the
optimal prior for precarious scene parameters will be a simple unimodal distribution with a single mean parameter
vector (and associated covariance matrix). (2) Rendering,
while readily expressed as a feed-forward function, is not
naturally differentiable at object boundaries (where small
changes in parameters can generate large changes in the
rendered image). While approximate differentiable renderers do exist [28], we wish to make use of highly-optimized
commercial packages for animation and image synthesis
(such as Unity 3D). As such, we adopt a simple samplingbased approach that addresses both limitations:
min
I
max
D
V ( D; I ) = [Imposter Selection] (3)
E
x s p t ( x ) [log D ( x )] + E z s Unif ( Z I ) [log(1 - D ( G ( z )))]
where Z I ⊆ Z . That is, we search for a subset of parameter
vectors (the “imposters”) that fool the discriminator. One
could employ various sequential sampling strategies for optimizing the above; start with a random sample of parameter vectors, update the discriminator (with gradient based
updates using a batch of real and synthesized data), generate additional samples close to those imposters that fool
the discriminator, and repeat. We found a single iteration to
work quite well. Our algorithm for synthesizing a realistic
set of precarious scenes is given in Alg. 1, and the overall
approach for advesarial domain adaption is given in Alg. 2.
Algorithm 1 Imposter Selection
Input: Set of examples from source domain S and target
domain T .
Output: Subset of imposters I ⊆ S .
1. Train a binary discriminator network D ( x ) that distinguishes examples x 2 S from x 2 T .
2. Return the subset of k samples from S that best fool
the discriminator.
Here, the set S consists of synthetic image-label pairs
rendered from an exhaustive set of scene parameters f z g
and the set T consists of real (Precarious) image-label pairs.
Without Step 2, Alg. 2 reduces to standard fine-tuning from
Algorithm 2 Domain Adaption with Imposters
Input: Set of examples from source domain S and target
domain T .
Output: Predictor f ( x ) for target set T .
1. Pre-Train a predictor f ( x ) on source set S .
2. Adapt the predictor on T [ I , where I is the set of
imposters found with Alg. 1.
3. Fine-tune the predictor on only target set T .
Figure 5: Architecture of RPN+.
a source to a target domain. Step 2 can be thought of as
“marginal distribution adaption”, since the distribution of
imposter images p I ( x ) mimics the true target distribution
p t ( x ) , at least from the discriminator’s perspective. But importantly, the discriminator D ( x ) has not made use of labels
y to find imposters, and so imposter labels may not mimic
the true target label distribution. Because of this, we opt to
finally fine-tune f ( x ) on the target image-label pairs. Alternatively, one may explore a discriminator that directly
operates on pairs of data and labels, as in [21].
4.1. Implementation
Discriminator D ( x ) : Our discriminator D is a VGG16
network trained to output the probability that an input images is real (with label 1) or synthetic (with label 0). We
found that modest number of images sufficed for training:
500 images from the Precarious Pedestrian train split and
1000 random synthetic images. We downsample images to
384 × 288 to accelerate training. After training D , we generate another set of 8000 synthetic images and select various
subsets of size k to define the imposter set (examined further in our experiments). We roughly find that 2.5% of the
synthesized images can serve as reasonable imposters.
Predictor f ( x ) : We make use of a detection system based
off on a region proposal network (RPN) [38, 49]. Rather
than training a RPN to return objectness proposals, we train
it to directly return pedestrian bounding boxes. Our network, denoted as RPN+, is illustrated in Figure 5. RPN+ is a
fully convolutional network implemented with TensorFlow.
We concatenate several layers on different stages in order
to improve the ability of locating people in different resolutions. We use 9 anchors (reference boxes with 3 scales
and aspect ratios) at each sliding position. During training,
a candidate bounding box will be treated as a positive if
its intersection-over-union overlap with a ground-truth box
exceeds 50%, and will be a negative for overlaps less than
20%. To accelerate training time, we initialize with a pretrained VGG-16 model where the first two convolutional
layers are frozen.
5. Experiments
5.1. Evaluation
We follow the evaluation protocol of the Caltech pedestrian dataset [10], which use ROC curves for 2D bounding
box detection at 50% and 70% overlap thresholds.
Testsets: We use three different datasets for evaluation:
our novel Precarious Pedestrian testset of real images, our
novel Adverserial Imposter Testset , and for diagnostics, a
standard pedestrian benchmark dataset ( Caltech ).
Baselines: We compare our approach with the following
baselines:
ACF: An aggregate channel features detector [8] .
LDCF: A LDCF detector [33].
HOG+Cascade: A cascade of boosted classifiers working
with HOG features [50].
HARR+Cascade: A cascade of boosted classifiers working with haar-like features[47, 25].
RPN/BF: A RPN detection model trained with boosted
forest [49], which appears to be the state-of-the-art pedestrian detection system at the time of publication.
Precarious Pedestrians: Results on Precarious Pedestrians
are presented in Figure 6. Our detector significantly outperforms alternative approaches, including the state-of-theart RPN/BF model. At 10 - 1 false positive per image, our
miss rate of 42.47% significantly outperforms all baselines,
including the state-of-the-art RPN/BF model (with a miss
rate of 54.5%). Note that all baseline detectors are trained
on Caltech. Comparing to baselines is complicated by the
fact that both the detection system and training dataset have
changed. However, in some sense, our fundamental contribution is method for generating more accurate training
datasets through adversarial imposters. To isolate the impact of our underlying detection network RPN+, we also
train a variant solely on the Caltech training set (denoted as
RPN+Caltech), making it directly comparable to all baselines because they use the same training set. RPN+Caltech
performs slightly worse than RPN/BF (with miss-rate of
58.82%), though it outperforms RPN/BF at higher false
positive rates. This suggests that our underlying network is
close to state-of-the-art, and moreover validates the signif-
detectors are trained on Caltech Dataset. For reference,
RPN+Caltech model would currently rank 6th out of 68 entries on the Caltech Dataset leaderboard. We also attempted
to evaluate our final model (trained with Adversarial Imposters) on Caltech, but saw lackluster performance. We
posit that this is due to the different set of scales and aspect
ratios in Precarious Pedestrians. We leave further crossdataset analysis to future work.
5.2. Diagnostics
In this section, we explore various variants of our approach. Table 2 examines different fine-tuning strategies
for adapting detectors from the source domain of synthetic
images to the target domain of real Precarious images. Finetuning via Imposters performs the best 42.47%, and no
Table 2: Miss rate of different fine-tuning strategies at a
false positive rate of 10 - 1 , where S , T , and I refer to source
datasets (of synthetic images), target dataset (of recarious
real images), and Imposter dataset.
ticeably outperforms the commonplace baselines of traditional fine-tuning (by 6%) and training on only the target
(by 24%).
Figure 10 examines the effect of k , the size of the imposter set. We find good performance when k is equal to
j T j , the size of the target set of Precarious Pedestrians used
for training. In retrospect, this may not be surprising as this
produces a balanced distribution of real images and Adversarial Imposters for training. Finally, Figure 10 also explores the impact of the discriminator. It plots performance
as a function of the training epoch used to learn D ( x ) . As
we train a better discriminator, the performance of our overall adversarial pipeline gets noticeably better.
6. Conclusion
We have explored methods for analyzing “in-the-tail”
urban scenes, which represent important modes of operations for autonomous vehicles. Motivated by the fact
that rare but dangerous scenes are exactly the scenarios
on which visual recognition should excel, we first analyze
existing datasets and illustrate that they do not contain
sufficient rare scenarios (because they naturally focus on
common or typical urban scenes). To address this gap, we
have collected our own dataset of Precarious Pedestrians,
which we will release to spur further research on this
important (but under explored) problem. Precarious scenes
are challenging because little data is available for both
evaluation and training. To address this challenge, we
propose the use of synthetic data generated with a game
engine. However, it is challenging to ensure that the
synthesized data matches the statistics of real precarious
scenarios. Inspired by generative adversarial networks, we
introduce the use of a discriminative classifier (trained to
discriminate real vs synthetic data) to implicitly specify this
distribution. We then use the synthesized data that fooled
the discriminator (the “synthetic imposters”) to both train
and evaluate state-of-the-art, robust pedestrian detection
systems.
Acknowledgements. This work was supported by NSF
(a) RPN+Caltech (b) Ours
Figure 7: Results on Precarious Dataset. The left shows results of RPN+ trained on Caltech, while the right shows RPN+
trained with adversarial imposters.
thetic dataset. ROC curves for detectors on Precarious
Dataset and Synthetic Dataset. Ours (RPN+) is trained with
selection model. The results suggest that performances of
algorithms on real dataset and synthetic dataset have the
same rankings.
Grant 1618903, NSF Grant 1208598, and Google. We
thank Yinpeng Dong and Mingsheng Long for helpful discussions.
References
[1] A. Agarwal and B. Triggs. A local basis representation for
estimating human pose from cluttered images. In Asian Conference on Computer Vision , pages 50–59. Springer, 2006.
[2] M. Andriluka, L. Pishchulin, P. Gehler, and B. Schiele. 2d
human pose estimation: New benchmark and state of the art
analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2014.
(a) Caltech Dataset (b) Caltech Dataset
Figure 9: (a) and (b) are ROC curves for different detectors under different overlap ratio criteria on Caltech Dataset
under the default “reasonable” test protocol.
[3] V. Athitsos, H. Wang, and A. Stefan. A database-based
framework for gesture recognition. Personal and Ubiquitous
Computing , 14(6):511–526, 2010.
[4] A. Broggi, A. Fascioli, P. Grisleri, T. Graf, and M. Meinecke.
Model-based validation approaches and matching techniques
for automotive vision based pedestrian detection. In 2005
IEEE Computer Society Conference on Computer Vision
and Pattern Recognition (CVPR’05)-Workshops , pages 1–1.
IEEE, 2005.
[5] X. Chen, Y. Duan, R. Houthooft, J. Schulman, I. Sutskever,
and P. Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets.
arXiv preprint arXiv:1606.03657 , 2016.
[6] N. Dalal and B. Triggs. Inria person dataset, 2005.
[7] E. L. Denton, S. Chintala, R. Fergus, et al. Deep generative image models using a laplacian pyramid of adversarial
networks. In Advances in neural information processing systems , pages 1486–1494, 2015.
[8] P. Dollar, R. Appel, S. Belongie, and P. Perona. Fast feature ´
pyramids for object detection. IEEE Transactions on Pattern
Analysis and Machine Intelligence , 36(8):1532–1545, 2014.
[9] P. Dollar, C. Wojek, B. Schiele, and P. Perona. Pedestrian ´
detection: A benchmark. In Computer Vision and Pattern
Recognition, 2009. CVPR 2009. IEEE Conference on , pages
304–311. IEEE, 2009.
[10] P. Dollar, C. Wojek, B. Schiele, and P. Perona. Pedestrian ´
detection: An evaluation of the state of the art. PAMI , 34,
2012.
[11] D. H. Eberly. 3D game engine design: a practical approach
to real-time computer graphics . CRC Press, 2006.
[12] M. Enzweiler and D. M. Gavrila. Monocular pedestrian detection: Survey and experiments. IEEE transactions on pattern analysis and machine intelligence , 31(12):2179–2195,
2009.
[13] A. Ess, B. Leibe, K. Schindler, , and L. van Gool. A mobile vision system for robust multi-person tracking. In IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR’08) . IEEE Press, June 2008.
[14] P. Fischer, A. Dosovitskiy, E. Ilg, P. Hausser, C. Hazırbas¸, ¨
V. Golkov, P. van der Smagt, D. Cremers, and T. Brox.
Flownet: Learning optical flow with convolutional networks.
arXiv preprint arXiv:1504.06852 , 2015.
[15] Y. Ganin and V. Lempitsky. Unsupervised domain adaptation
by backpropagation. arXiv preprint arXiv:1409.7495 , 2014.
[16] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle,
F. Laviolette, M. Marchand, and V. Lempitsky. Domainadversarial training of neural networks. Journal of Machine
Learning Research , 17(59):1–35, 2016.
[17] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu,
D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in Neural Information
Processing Systems , pages 2672–2680, 2014.
[18] K. Grauman, G. Shakhnarovich, and T. Darrell. Inferring
3d structure with a statistical image-based shape model. In
Computer Vision, 2003. Proceedings. Ninth IEEE International Conference on , pages 641–647. IEEE, 2003.
[19] H. Hattori, V. Naresh Boddeti, K. M. Kitani, and T. Kanade.
Learning scene-specific pedestrian detectors without real
data. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition , pages 3819–3827, 2015.
[20] M. Hejrati and D. Ramanan. Analysis by synthesis: 3d object recognition by object reconstruction. In 2014 IEEE Conference on Computer Vision and Pattern Recognition , pages
2449–2456. IEEE, 2014.
[21] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Imageto-image translation with conditional adversarial networks.
arXiv preprint arXiv:1611.07004 , 2016.
[22] D. P. Kingma and M. Welling. Auto-encoding variational
bayes. arXiv preprint arXiv:1312.6114 , 2013.
[23] K. Lai, L. Bo, and D. Fox. Unsupervised feature learning for
3d scene labeling. In 2014 IEEE International Conference on
Robotics and Automation (ICRA) , pages 3050–3057. IEEE,
2014.
[24] A. Lerer, S. Gross, and R. Fergus. Learning physical
intuition of block towers by example. arXiv preprint
arXiv:1603.01312 , 2016.
[25] R. Lienhart and J. Maydt. An extended set of haar-like features for rapid object detection. In Image Processing. 2002.
Proceedings. 2002 International Conference on , volume 1,
pages I–900. IEEE, 2002.
[26] M.-Y. Liu and O. Tuzel. Coupled generative adversarial networks. In Advances in Neural Information Processing Systems , pages 469–477, 2016.
[27] M. Long, Y. Cao, J. Wang, and M. I. Jordan. Learning transferable features with deep adaptation networks. In ICML ,
pages 97–105, 2015.
[28] M. M. Loper and M. J. Black. Opendr: An approximate differentiable renderer. In European Conference on Computer
Vision , pages 154–169. Springer, 2014.
[29] J. Marin, D. Vazquez, D. Ger ´ onimo, and A. M. L ´ opez. ´
Learning appearance in virtual scenarios for pedestrian
detection. In Computer Vision and Pattern Recognition
(CVPR), 2010 IEEE Conference on , pages 137–144. IEEE,
2010.
[30] N. Mayer, E. Ilg, P. Hausser, P. Fischer, D. Cremers, ¨
A. Dosovitskiy, and T. Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow
estimation. arXiv preprint arXiv:1512.02134 , 2015.
[31] M. Mirza and S. Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784 , 2014.
[32] Y. Movshovitz-Attias, Y. Sheikh, V. N. Boddeti, and Z. Wei.
3d pose-by-detection of vehicles via discriminatively reduced ensembles of correlation filters. In BMVC , 2014.
[33] W. Nam, P. Dollar, and J. H. Han. Local decorrelation for ´
improved pedestrian detection. In Advances in Neural Information Processing Systems , pages 424–432, 2014.
[34] B. Pepik, M. Stark, P. Gehler, and B. Schiele. Teaching 3d
geometry to deformable part models. In Computer Vision
and Pattern Recognition (CVPR), 2012 IEEE Conference on ,
pages 3362–3369. IEEE, 2012.
[35] L. Pishchulin, A. Jain, C. Wojek, M. Andriluka,
T. Thormahlen, and B. Schiele. Learning people detec- ¨
tion models from few training samples. In Computer Vision
and Pattern Recognition (CVPR), 2011 IEEE Conference on ,
pages 1473–1480. IEEE, 2011.
[36] M. Potamias and V. Athitsos. Nearest neighbor search methods for handshape recognition. In Proceedings of the 1st
international conference on PErvasive Technologies Related
to Assistive Environments , page 30. ACM, 2008.
[37] A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 , 2015.
[38] S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards
real-time object detection with region proposal networks. In
Advances in neural information processing systems , pages
91–99, 2015.
[39] D. J. Rezende, S. Mohamed, and D. Wierstra. Stochastic
backpropagation and approximate inference in deep generative models. arXiv preprint arXiv:1401.4082 , 2014.
[40] S. R. Richter, V. Vineet, S. Roth, and V. Koltun. Playing
for data: Ground truth from computer games. arXiv preprint
arXiv:1608.02192 , 2016.
[41] J. Romero, H. Kjellstrom, and D. Kragic. Hands in action: ¨
real-time 3d reconstruction of hands in interaction with objects. In Robotics and Automation (ICRA), 2010 IEEE International Conference on , pages 458–463. IEEE, 2010.
[42] G. Ros, L. Sellart, J. Materzynska, D. Vazquez, and A. M.
Lopez. The synthia dataset: A large collection of synthetic
images for semantic segmentation of urban scenes. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition , pages 3234–3243, 2016.
[43] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,
S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,
et al. Imagenet large scale visual recognition challenge.
International Journal of Computer Vision , 115(3):211–252,
2015.
[44] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen. Improved techniques for training gans.
arXiv preprint arXiv:1606.03498 , 2016.
[45] S. Satkin, J. Lin, and M. Hebert. Data-driven scene understanding from 3d models. 2012.
[46] B. Sun and K. Saenko. From virtual to reality: Fast adaptation of virtual object detectors to real domains. In BMVC ,
volume 1, page 3, 2014.
[47] P. Viola and M. Jones. Rapid object detection using a boosted
cascade of simple features. In Computer Vision and Pattern
Recognition, 2001. CVPR 2001. Proceedings of the 2001
IEEE Computer Society Conference on , volume 1, pages I–
511. IEEE, 2001.
[48] C. Wojek, S. Walk, and B. Schiele. Multi-cue onboard pedestrian detection. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on , pages 794–
801. IEEE, 2009.
[49] L. Zhang, L. Lin, X. Liang, and K. He. Is faster r-cnn doing
well for pedestrian detection? In European Conference on
Computer Vision , pages 443–457. Springer, 2016.
[50] Q. Zhu, M.-C. Yeh, K.-T. Cheng, and S. Avidan. Fast human
detection using a cascade of histograms of oriented gradients. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06) , volume 2,
pages 1491–1498. IEEE, 2006.
Shiyu Huang
Tsinghua University
Beijing, China
[email protected]
Deva Ramanan
Carnegie Mellon University
Pittsburgh, USA
[email protected]
Abstract
As autonomous vehicles become an every-day reality,
high-accuracy pedestrian detection is of paramount practical importance. Pedestrian detection is a highly researched
topic with mature methods, but most datasets focus on common scenes of people engaged in typical walking poses on
sidewalks. But performance is most crucial for dangerous
scenarios, such as children playing in the street or people
using bicycles/skateboards in unexpected ways. Such “inthe-tail” data is notoriously hard to observe, making both
training and testing difficult. To analyze this problem, we
have collected a novel annotated dataset of dangerous scenarios called the Precarious Pedestrian dataset.
Even given a dedicated collection effort, it is relatively
small by contemporary standards ( ≈ 1000 images). To allow for large-scale data-driven learning, we explore the use
of synthetic data generated by a game engine. A significant
challenge is selected the right “priors” or parameters for
synthesis: we would like realistic data with poses and object configurations that mimic true Precarious Pedestrians.
Inspired by Generative Adversarial Networks (GANs), we
generate a massive amount of synthetic data and train a discriminative classifier to select a realistic subset, which we
deem the Adversarial Imposters. We demonstrate that this
simple pipeline allows one to synthesize realistic training
data by making use of rendering/animation engines within
a GAN framework. Interestingly, we also demonstrate that
such data can be used to rank algorithms, suggesting that
Adversarial Imposters can also be used for “in-the-tail”
validation at test-time, a notoriously difficult challenge for
real-world deployment.
(a) (b) (c)
Figure 1: (a) Examples from our novel Precarious Pedestrian Dataset of dangerous, but rare pedestrian scenes. One
important scenario is that of pedestrians on their phone (row
2, col 2), who may not be adequately aware of their surroundings. (b) Examples in Caltech Dataset tend not to
capture such rare scenarios. (c) Examples from a set of
Adveserial Imposters, which are synthetic images that are
adversarially-trained to mimic the set of Precarious Pedstrians. We demonstrate that such images can be used to both
train and evaluate robust pedestrian recognition systems targeting such dangerous scenarios.
1. Introduction
There’s no software designer in the world that’s ever going to be
smart enough to anticipate all the potential circumstances an
autonomous car is going to encounter. The dog that runs out into
the street, the person who runs up the street, the bicyclist, the
policeman or the construction worker.
C. Hart, Chairman of National Transport. Safety Board
As autonomous vehicles become an every-day reality,
high-accuracy pedestrian detection is of paramount practical importance. Pedestrian detection is a highly researched
arXiv:1703.06283v2 [cs.CV] 10 Apr 2017
topic with mature methods, but most datasets focus on
“everyday” scenes of people engaged in typical walking
poses on sidewalks [9, 6, 13, 12, 48]. However, perhaps
the most important operating point for a deployable system is its behaviour in dangerous, unexpected scenarios,
such as children playing in the street or people using bicycles/skateboards in unexpected ways.
Precarious Pedestrian Dataset: Such “in-the-tail” data is
notoriously hard to observe, making both training and evaluation of existing systems difficult. To analyze this problem, we have collected a novel annotated dataset of dangerous scenarios called the Precarious Pedestrian Dataset.
Even given a dedicated collection effort, it is relatively
small by contemporary standards ( ≈ 1000 images). To explore large-scale data-driven learning, we explore the use of
synthetic data generated by a game engine. Synthetic training data is an actively explored topic because it provides a
potentially infinite well of annotated data for training datahungry architectures [24, 30, 14, 19, 40, 42]. Particularly
attractive are approaches that combine a large amount of
synthetic training data with a small amount of real data (that
may have been difficult to acquire and/or label).
Challenges in Synthesis: We see two primary difficulties
with the use of synthetic training data. The first is that not
all data is created “equal”: when combining synthetic data
with real data, synthesizing common scenes may not be particularly useful since they will likely already appear in the
training set. Hence we argue that the real power of synthetic
data is generating examples “in-the-tail”, which would otherwise have been hard to collect. The second difficulty
arises in building good generative models of images, a notoriously difficult problem. Rather than building generative
pixel-level models, we make use of state-of-the-art rendering/animation engines that contain an immense amount of
knowledge (about physics, light transfer, etc.). The challenge of generative synthesis then lies in constructing the
right “priors”, or scene-parameters, to render/animate. In
our case, these correspond to body poses and spatial configurations of people and other objects in the scene.
Adversarial Imposters: We address both concerns
with a novel variant of Generative Adversarial Networks
(GANs) [17], a method for synthesizing data from latent
noise vectors. Traditional GANs learn generative feedforward models that process latent noise vectors, typically
from a fixed known prior distribution. Instead, we fix the
feedforward model to be a rendering engine, but use an adverserial framework to learn the latent priors. To do so,
we define a rendering pipeline that takes an input a vector of scene parameters capturing object attributes and spatial layout. We use rejection sampling to construct a set
of scene parameters (and their associated rendered images)
that maximally confuse the discriminator. We call such examples Adversarial Imposters, and use them within a sim-
(a) (b)
Figure 2: (a) Scene that’s built for generating synthetic images. (b) 3D models that we use in this project.
ple pipeline for adapting detectors from synthetic data to the
world of real images.
RPN+: We use our dataset of real and imposter images to
train a suite of contemporary detectors. We find surprisingly good results with a (to our knowledge) novel variant
of region proposal network (RPN) [49] tuned for particular objects (precarious people) rather than a general class of
objectness detections. Instead of classifying a sparse set of
proposed windows (as nearly all contemporary object detection systems based on RCNN do [38]), this network returns a dense heatmap of pedestrian detections, along with
regressed bounding box location for each pixel location in
the heatmap. We call this detector RPN+. Our experiments
show that our RPN+, trained on real+imposter data, outperforms other detectors trained only on real data.
Validation: Interestingly, we also demonstrate that our Adverserial Imposter Dataset can be used to rank algorithms,
suggesting that our pipeline can also be used for “in-thetail” validation at test-time, a notoriously difficult challenge
for real-world deployment.
Contributions: The contribution of our work is as follows:
(1) a novel dataset of pedestrians in dangerous situations
(Precarious Pedestrians) (2) a general architecture for creating realistic synthetic data “in-the-tail”, for which limited real data can be collected and (3) demonstration of our
overall pipeline for the task of pedestrian detection using a
novel detector. Our datasets and code can be found here:
https://github.com/huangshiyu13/RPNplus.
2. Related work
Synthetic Data: Synthetic datasets have been used to train
and evaluate the performance of computer vision algorithms. Some forms of ground truth are hard to obtain from
hand-labelling, such as optical flow, but easy to synthesize
via simulation [14]. Adam et al . [24] used a 3D game
engine to generate synthetic data and learned an intuitive
physics model to predict falling towers of blocks. Mayer et
al . [30] released a benchmark suite of various tasks using
synthetic data, including disparity and optical flow. Richter
et al . [40] used synthetic data to improve image segmentation performance, but notably do not control the scene as
to explore targeted arrangements of objects. German Ros
et al . [42] used Unity Development Platform to generate a
synthetic urban scene dataset.
3D Models for Detection: A notable application of 3D
computer graphics model in vision has been the modeling of the human body shapes [18, 4, 1, 29, 36, 3, 41].
Moreover, 3D simulation can also been used for car detection [34, 32, 20] and scene understanding [45, 23]. Marin et
al . [29] used a game engine to generate synthetic training
data. Pishchulin et al . [35] used 8 HD cameras to scan human body and built real 3D human models. Then they used
synthetic data and some labelled real data to train pedestrian
detectors. Hattori et al . [19] used 3D modelling software
to build a special scene and randomly put 3D models on a
special background for pedestrian detection. Most of these
works use synthetic data “as-is”, while we analyze statistical differences between synthetic and real data, describing a
pipeline for reconciling such differences through adversarial domain adaption.
Domain Adaptation: Domain Adaptation is a standard
strategy to deal with data across different domains, such as
synthetic versus real. Large synthetic datasets can be used
to bootstrap detectors and then adapted to real data by moving to the target domain distribution. Sun and Saenko [46]
used 3D models to train detectors for real objects. Such
work typically used shallow detectors defined on fixed feature sets, while we focus on gradient-based adaption of
“deep” detection networks (such as RCNN). From this perspective, our work is inspired by approaches for deep domain adaptation [15, 16, 26, 27]. Such work typically assumes that one has access to large amounts of unlabeled
data from the target domain. In our case, assembling a large
target dataset of unlabelled examples (of real Precarious
Pedestrians) is itself challenging, necessitating the need for
alternative approaches that make stronger use of the source
dataset.
Generative Adversarial Nets: GANs [17] are deep networks that can generate synthetic images from latent noise
vectors. They do so by adversarially-training a neural network to discriminate between real versus synthetic images.
Recent works have shown impressive performance in generation of synthetic images [31, 7, 37, 44, 5]. However,
it appears challenging to synthesize high-resolution images
with semantically-valid content. We circumvent these limitations with a rendering-based adversarial approach to image synthesis.
(a) Precarious Dataset (b) Caltech Dataset
| (c) Precarious Dataset | (d) Caltech Dataset |
| Figure 3: (a) and (b) show the percentage of the number | |
| of people per image in both datasets. (c) and (d) show the | |
| percentage of the different types of people in both datasets. | |
| Precarious Dataset contains more cyclists and motorcyclists | |
| than Caltech Dataset. | |
| 3. Datasets | |
| 3.1. Precarious Pedestrian Dataset | |
| We begin by describing our Precarious Pedestrian | |
| Dataset. We perform a dedicated search for targeted key | |
| words (such as “pedestrian fall”, “traffic violation” and | |
| “dangerous bike rider”) on Google Images, Baidu Images, | |
| and some selected images from MPII Dataset [2], producing | |
| a total of 951 reasonable images. We then label bounding | |
| boxes for each image manually. Precarious Pedestrians con | |
| tains various kinds scenes, such as children running on the | |
| road, people tripping, motorcyclists performing dangerous | |
| movements, people interacting with objects (such as bicy | |
| cles or umbrellas). One important dangerous but increas | |
| ingly common scenario consists of people watching their | |
| phones or texting while crossing the street, which is poten | |
| tially dangerous as the person may not be aware of their | |
| surroundings (Figure 1). To quantify the (dis)similarity of | |
| Precarious Pedestrians to standard pedestrian benchmarks | |
| such as Caltech [10], we tabulate the percentage of images | |
| with more than one people, as well as the number of irregu | |
| lar “pedestrians” such as bicyclists or motorcyclists. Com | |
| pared to Caltech, Precarious Pedestrians contains images | |
| with many more overall people as well as many more cy | |
| clists and motorbikes (Figure 3). We split the Precarious | |
| dataset equally for training and testing. | |
| 3.2. Synthetic Dataset | |
| To help both train and evaluate algorithms for detecting | |
| precarious pedestrians, we make use of a synthetic data. In | |
| this section, we describe our rendering pipeline for gener | |
| ating synthetic data. We use the Unity 3D game engine as | |
| our basic platform for simulation and rendering, due to the | |
| large availability of both commercial and user-generated | |
| assets, in the form of 3D models and character animations. |
| Range | |
| Number of 3D models Index of background images Index of 3D models Position of 3D models Index of Animations Time of animation Model’s angle on the x axis Model’s angle on the y axis Model’s angle on the z axis Light intensity Light’s angle on the x axis Light’s angle on the y axis |
[4; 8] [0; 1726) [0; 20) Within the field of vision [0; maxnumber) [0; 1] [-90◦; 90◦] [-180◦; 180◦] [-90◦; 90◦] [0:5; 2] [-45◦; 45◦] [-45◦; 45◦] |
Table 1: Constraints of parameters for synthesizing images.
The index and time(normalized) of animations will jointly
decide the gestures of 3D models .
Figure 2 shows the commercial 3D human models
that we use for data generation, consisting of 20 models
spanning different women, men, cyclists and skateboarder
avatars. Because these are designed for game engine
play, each 3D model is associated with characteristic
animations such as jumping, talking, running, cheering and
applauding. We animate these models in a 3D scene with
a 2D billboard to capture the scene background [11], as
shown in Figure 2. Billboards are randomly sampled from
the 1726 background images from INRIA dataset [6] and
a custom set of outdoor scenes downloaded from Internet.
Our approach can generate a diverse set of background
scenes, unlike approaches that are limited to a single virtual
urban city [29].
Scene parameters: To build a large library of synthetic images that will potentially be used for training and evaluation,
we first define a set of parameters and parameter ranges. We
index the set of background images, the set of 3D models,
and the animation frame number for each model. In brief,
the scene parameters include directional light intensity and
direction (capturing sunlight), the background image index,
the number of 3D models, and for each model, an index
specifying the avatar ID and animation frame, as well as a
root position and orientation (rotation in the ground plane).
We assume a fixed camera viewpoint. Note that the root position affects both the location and scale of the 3D model
in the rendered image. All these parameters can be summarized as a variable-length vector z 2 Z , where each vector
corresponds to a particular scene instantiation.
Synthesis: Our generator G ( z ) , or rendering engine, synthesizes an image corresponding to z . Importantly, we can
also synthesize labels L ( z ) for each rendered image, spec-
(a) (b)
Figure 4: (a) Imposter images that are chosen by selector.
(b) Synthetic images that are not in Imposter Dataset.
ifying object type, 3D location, pixel segmentation masks,
etc . In practice, we make use of only 2D object bounding
boxes. Table 1 shows the viable ranges of each parameter.
In addition, we found the following heuristic to simulate
reasonable object layouts: we enforce the maximum overlap between any two 3D models to be 20% (to avoid congestion) and the projected location of the 3D models should
lie within the camera’s field-of-view. These conditions are
straightforward to verify for a given vector z without rendering any pixels, and so can be efficiently enforced though
rejection sampling (i.e., generate a random vector and only
render those that pass these conditions). Unlike Hironori et
al . [19], who generate training data by manually tuning z
to match specific scenes, our approach is not scene specific
and does not require any manual intervention.
Pre-processing: Synthesized images and Precarious Pedestrian images may be of different sizes. We isotropically
scale each image to a resolution of 960 × 720, zero-padding
as necessary. Our experiments also make use of the Caltech Pedestrian benchmark, to which we apply the same
pre-processing.
4. Proposed Method
Domain adaption: In this section, we introduce a novel
framework for adversarially adapting detectors from synthetic training data to real training data. We use x 2 X to
denote an image and y 2 Y to denote its label vector (a set
of bounding box labels). Let p s ( x ; y ) to refer to the distribution of image-label pairs from the source domain (of
synthetic images), and p t ( x ; y ) to refer to the target domain
(of real Precarious Pedestrians). In our problem, we expect large amounts of source samples, but a limited amount
of target ones. We factorize the joint into a marginal over
image appearance and conditional on label given the appearance - e.g., p s ( x ) p s ( y j x ) . Importantly, we discriminatively train a feedforward function f s ( x ) = p s ( y j x ) to
match the conditional distribution. Our central question is
how to transfer feedforward predictors trained from source
samples f s ( x ) to the target domain f t ( x ) .
Fine-tuning: The most natural approach to domain adaption may simply be to fine-tune a predictor f s ( x ) , originally trained on the source, with samples from the target
p t ( x ; y ) . Indeed, virtually all contemporary methods for visual recognition makes use of fine-tuned models that were
pre-trained on Imagenet [43]. We compare to such a strategy in our experiments, but find that fine-tuning works best
when source and target distributions are similar. As we
argue, while rendering engines can produce photorealistic
scenes, it is difficult to specify a prior over scene parameters
that mimic real (Precarious) scenes. We describe a solution
that adversarially learns a prior.
Generators: As introduced in Sec. 3.2, let z 2 Z be a
vector of scene parameters, G ( z ) 2 X be a feedforward
generator function that renders a synthetic image given the
scene parameters, and L ( z ) 2 Y be a function that generates labels from the scene parameters. We can then reparameterize the distribution over synthetic images as a distribution over scene parameters p z ( z ) . We now describe a
procedure for learning a prior p z ( z ) that allows for easier
transfer. Specifically, we learn a prior that fools an adversary that is trying to distinguish samples from the source
and target.
Adversarial generators: To describe our approach, we first
recall a traditional generative adverserial network (GAN):
min
G
max
D
V ( D; G ) = [Gen. Adversarial Net] (1)
E
x s p t ( x ) [log D ( x )] + E z s p z ( z ) [log(1 - D ( G ( z )))]
where the minmax optimization jointly tries to estimate a
discriminator D that can distinguish real versus synthesized
data examples, and the generator G tries to synthesize realistic examples that fool the discriminator. Typically, the
discriminator D ( x ) is trained to output the probability that
x is real (e.g., a real Precarious Pedestrian), while p z ( z )
is fixed to be a zero-mean, unit-variance Gaussian. This
optimization can be performed with stochastic gradient updates, that converge (in the limit) to a fixed point of the minimax problem. We refer the reader to the excellent introduction in [17]. Importantly, the generator must encode complex constraints about the manifold of natural images, that
capture amongst other knowledge the physical properties of
light transport and material appearance.
Adversarial priors: We note that rendering engines can
be viewed as generators that already contain much of this
knowledge, and so we fix G to be a production-quality rendering platform (Unity 3D). Instead, we learn the prior over
parameter vectors in a adversarial manner:
min
I
max
D
V ( D; I ) = [Adversarial Priors] (2)
E
x s p t ( x ) [log D ( x )] + E z s p I ( z ) [log(1 - D ( G ( z )))]
If the generator G is differentiable with respect to z , it is
possible to use backprop to compute gradient updates for
simple prior distributions p I ( z ) , such as Gaussians [22, 39].
This implies that the above formulation of adversarial priors
is amenable to gradient-based learning.
Imposter search: We see two difficulties with directly applying (2) to our problem: (1) It seems unlikely that the
optimal prior for precarious scene parameters will be a simple unimodal distribution with a single mean parameter
vector (and associated covariance matrix). (2) Rendering,
while readily expressed as a feed-forward function, is not
naturally differentiable at object boundaries (where small
changes in parameters can generate large changes in the
rendered image). While approximate differentiable renderers do exist [28], we wish to make use of highly-optimized
commercial packages for animation and image synthesis
(such as Unity 3D). As such, we adopt a simple samplingbased approach that addresses both limitations:
min
I
max
D
V ( D; I ) = [Imposter Selection] (3)
E
x s p t ( x ) [log D ( x )] + E z s Unif ( Z I ) [log(1 - D ( G ( z )))]
where Z I ⊆ Z . That is, we search for a subset of parameter
vectors (the “imposters”) that fool the discriminator. One
could employ various sequential sampling strategies for optimizing the above; start with a random sample of parameter vectors, update the discriminator (with gradient based
updates using a batch of real and synthesized data), generate additional samples close to those imposters that fool
the discriminator, and repeat. We found a single iteration to
work quite well. Our algorithm for synthesizing a realistic
set of precarious scenes is given in Alg. 1, and the overall
approach for advesarial domain adaption is given in Alg. 2.
Algorithm 1 Imposter Selection
Input: Set of examples from source domain S and target
domain T .
Output: Subset of imposters I ⊆ S .
1. Train a binary discriminator network D ( x ) that distinguishes examples x 2 S from x 2 T .
2. Return the subset of k samples from S that best fool
the discriminator.
Here, the set S consists of synthetic image-label pairs
rendered from an exhaustive set of scene parameters f z g
and the set T consists of real (Precarious) image-label pairs.
Without Step 2, Alg. 2 reduces to standard fine-tuning from
Algorithm 2 Domain Adaption with Imposters
Input: Set of examples from source domain S and target
domain T .
Output: Predictor f ( x ) for target set T .
1. Pre-Train a predictor f ( x ) on source set S .
2. Adapt the predictor on T [ I , where I is the set of
imposters found with Alg. 1.
3. Fine-tune the predictor on only target set T .
Figure 5: Architecture of RPN+.
a source to a target domain. Step 2 can be thought of as
“marginal distribution adaption”, since the distribution of
imposter images p I ( x ) mimics the true target distribution
p t ( x ) , at least from the discriminator’s perspective. But importantly, the discriminator D ( x ) has not made use of labels
y to find imposters, and so imposter labels may not mimic
the true target label distribution. Because of this, we opt to
finally fine-tune f ( x ) on the target image-label pairs. Alternatively, one may explore a discriminator that directly
operates on pairs of data and labels, as in [21].
4.1. Implementation
Discriminator D ( x ) : Our discriminator D is a VGG16
network trained to output the probability that an input images is real (with label 1) or synthetic (with label 0). We
found that modest number of images sufficed for training:
500 images from the Precarious Pedestrian train split and
1000 random synthetic images. We downsample images to
384 × 288 to accelerate training. After training D , we generate another set of 8000 synthetic images and select various
subsets of size k to define the imposter set (examined further in our experiments). We roughly find that 2.5% of the
synthesized images can serve as reasonable imposters.
Predictor f ( x ) : We make use of a detection system based
off on a region proposal network (RPN) [38, 49]. Rather
than training a RPN to return objectness proposals, we train
it to directly return pedestrian bounding boxes. Our network, denoted as RPN+, is illustrated in Figure 5. RPN+ is a
fully convolutional network implemented with TensorFlow.
We concatenate several layers on different stages in order
to improve the ability of locating people in different resolutions. We use 9 anchors (reference boxes with 3 scales
and aspect ratios) at each sliding position. During training,
a candidate bounding box will be treated as a positive if
its intersection-over-union overlap with a ground-truth box
exceeds 50%, and will be a negative for overlaps less than
20%. To accelerate training time, we initialize with a pretrained VGG-16 model where the first two convolutional
layers are frozen.
5. Experiments
5.1. Evaluation
We follow the evaluation protocol of the Caltech pedestrian dataset [10], which use ROC curves for 2D bounding
box detection at 50% and 70% overlap thresholds.
Testsets: We use three different datasets for evaluation:
our novel Precarious Pedestrian testset of real images, our
novel Adverserial Imposter Testset , and for diagnostics, a
standard pedestrian benchmark dataset ( Caltech ).
Baselines: We compare our approach with the following
baselines:
ACF: An aggregate channel features detector [8] .
LDCF: A LDCF detector [33].
HOG+Cascade: A cascade of boosted classifiers working
with HOG features [50].
HARR+Cascade: A cascade of boosted classifiers working with haar-like features[47, 25].
RPN/BF: A RPN detection model trained with boosted
forest [49], which appears to be the state-of-the-art pedestrian detection system at the time of publication.
Precarious Pedestrians: Results on Precarious Pedestrians
are presented in Figure 6. Our detector significantly outperforms alternative approaches, including the state-of-theart RPN/BF model. At 10 - 1 false positive per image, our
miss rate of 42.47% significantly outperforms all baselines,
including the state-of-the-art RPN/BF model (with a miss
rate of 54.5%). Note that all baseline detectors are trained
on Caltech. Comparing to baselines is complicated by the
fact that both the detection system and training dataset have
changed. However, in some sense, our fundamental contribution is method for generating more accurate training
datasets through adversarial imposters. To isolate the impact of our underlying detection network RPN+, we also
train a variant solely on the Caltech training set (denoted as
RPN+Caltech), making it directly comparable to all baselines because they use the same training set. RPN+Caltech
performs slightly worse than RPN/BF (with miss-rate of
58.82%), though it outperforms RPN/BF at higher false
positive rates. This suggests that our underlying network is
close to state-of-the-art, and moreover validates the signif-
| (a) Precarious Dataset | (b) Precarious Dataset |
| Figure 6: (a) and (b) are ROC curves for different detec | |
| tors under different overlap ratio criteria on the Precarious | |
| Pedestrian testset. In the legend, we denote the miss rate at | |
| 10-1 false positives per image. RPN+Caltech refers to our | |
| RPN+ network architecture trained only on Caltech, while | |
| Ours refers to our detector (RPN+) trained on synthetic, im | |
| poster, and real images (Alg 2). Note all detectors besides | |
| Ours are trained on the Caltech Dataset. | |
| icant improvement of training with Adversarial Imposters. | |
| Figure 7 visualizes the results of RPN+, both trained on Cal | |
| tech and trained with Adversarial Imposters. Qualitatively, | |
| we find that Precarious Pedestrians tend to take on more | |
| pose variation than typical pedestrians. This requires detec | |
| tion systems that are able to report back a wider range of | |
| bounding box scales and aspect ratios. | |
| Adversarial Imposters: We also explore how detectors | |
| perform on a testset of Adversarial Imposters. Note that | |
| we can generate an arbitrarily large testset since it is syn | |
| thetic. Figure 8 and Figure 10 show that the performance | |
| on both real test data and synthetic test data has the same | |
| ranking order. These results suggest that synthetic data may | |
| be useful as a testset for evaluating detectors on rare (but | |
| important) scenarios that are difficult to observe in real test | |
| data. | |
| Caltech: Finally, for completeness, we also test our RPN+ | |
| network on the Caltech Dataset in Figure 9. Here, all the |
detectors are trained on Caltech Dataset. For reference,
RPN+Caltech model would currently rank 6th out of 68 entries on the Caltech Dataset leaderboard. We also attempted
to evaluate our final model (trained with Adversarial Imposters) on Caltech, but saw lackluster performance. We
posit that this is due to the different set of scales and aspect
ratios in Precarious Pedestrians. We leave further crossdataset analysis to future work.
5.2. Diagnostics
In this section, we explore various variants of our approach. Table 2 examines different fine-tuning strategies
for adapting detectors from the source domain of synthetic
images to the target domain of real Precarious images. Finetuning via Imposters performs the best 42.47%, and no
| Fine-tuning method | 50% overlap | 70% overlap |
| S T S ) T S ) (T [ I) S ) (T [ I) ) T |
83.49% 72.39% 48.45% 45.97% 42.47% |
95.18% 93.70% 77.14% 74.94% 73.70% |
Table 2: Miss rate of different fine-tuning strategies at a
false positive rate of 10 - 1 , where S , T , and I refer to source
datasets (of synthetic images), target dataset (of recarious
real images), and Imposter dataset.
ticeably outperforms the commonplace baselines of traditional fine-tuning (by 6%) and training on only the target
(by 24%).
Figure 10 examines the effect of k , the size of the imposter set. We find good performance when k is equal to
j T j , the size of the target set of Precarious Pedestrians used
for training. In retrospect, this may not be surprising as this
produces a balanced distribution of real images and Adversarial Imposters for training. Finally, Figure 10 also explores the impact of the discriminator. It plots performance
as a function of the training epoch used to learn D ( x ) . As
we train a better discriminator, the performance of our overall adversarial pipeline gets noticeably better.
6. Conclusion
We have explored methods for analyzing “in-the-tail”
urban scenes, which represent important modes of operations for autonomous vehicles. Motivated by the fact
that rare but dangerous scenes are exactly the scenarios
on which visual recognition should excel, we first analyze
existing datasets and illustrate that they do not contain
sufficient rare scenarios (because they naturally focus on
common or typical urban scenes). To address this gap, we
have collected our own dataset of Precarious Pedestrians,
which we will release to spur further research on this
important (but under explored) problem. Precarious scenes
are challenging because little data is available for both
evaluation and training. To address this challenge, we
propose the use of synthetic data generated with a game
engine. However, it is challenging to ensure that the
synthesized data matches the statistics of real precarious
scenarios. Inspired by generative adversarial networks, we
introduce the use of a discriminative classifier (trained to
discriminate real vs synthetic data) to implicitly specify this
distribution. We then use the synthesized data that fooled
the discriminator (the “synthetic imposters”) to both train
and evaluate state-of-the-art, robust pedestrian detection
systems.
Acknowledgements. This work was supported by NSF
(a) RPN+Caltech (b) Ours
Figure 7: Results on Precarious Dataset. The left shows results of RPN+ trained on Caltech, while the right shows RPN+
trained with adversarial imposters.
| (a) Precarious Dataset | (b) Synthetic Dataset |
| Figure 8: Algorithm rankings on real datasets and syn |
thetic dataset. ROC curves for detectors on Precarious
Dataset and Synthetic Dataset. Ours (RPN+) is trained with
selection model. The results suggest that performances of
algorithms on real dataset and synthetic dataset have the
same rankings.
Grant 1618903, NSF Grant 1208598, and Google. We
thank Yinpeng Dong and Mingsheng Long for helpful discussions.
References
[1] A. Agarwal and B. Triggs. A local basis representation for
estimating human pose from cluttered images. In Asian Conference on Computer Vision , pages 50–59. Springer, 2006.
[2] M. Andriluka, L. Pishchulin, P. Gehler, and B. Schiele. 2d
human pose estimation: New benchmark and state of the art
analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2014.
(a) Caltech Dataset (b) Caltech Dataset
Figure 9: (a) and (b) are ROC curves for different detectors under different overlap ratio criteria on Caltech Dataset
under the default “reasonable” test protocol.
| (a) Number of imposters | (b) Number of epochs |
| Figure 10: (a) Results of different amounts of Imposters. | |
| Optimal performance is obtained with an imposter set that is | |
| roughly equal in size to the set of 500 target samples of Pre | |
| carious training images. (b) Results of different selectors. | |
| We choose three selectors from different training periods | |
| and use them to select imposters separately. As the discrim | |
| inator gets better, so does the final fine-tuned detector. |
[3] V. Athitsos, H. Wang, and A. Stefan. A database-based
framework for gesture recognition. Personal and Ubiquitous
Computing , 14(6):511–526, 2010.
[4] A. Broggi, A. Fascioli, P. Grisleri, T. Graf, and M. Meinecke.
Model-based validation approaches and matching techniques
for automotive vision based pedestrian detection. In 2005
IEEE Computer Society Conference on Computer Vision
and Pattern Recognition (CVPR’05)-Workshops , pages 1–1.
IEEE, 2005.
[5] X. Chen, Y. Duan, R. Houthooft, J. Schulman, I. Sutskever,
and P. Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets.
arXiv preprint arXiv:1606.03657 , 2016.
[6] N. Dalal and B. Triggs. Inria person dataset, 2005.
[7] E. L. Denton, S. Chintala, R. Fergus, et al. Deep generative image models using a laplacian pyramid of adversarial
networks. In Advances in neural information processing systems , pages 1486–1494, 2015.
[8] P. Dollar, R. Appel, S. Belongie, and P. Perona. Fast feature ´
pyramids for object detection. IEEE Transactions on Pattern
Analysis and Machine Intelligence , 36(8):1532–1545, 2014.
[9] P. Dollar, C. Wojek, B. Schiele, and P. Perona. Pedestrian ´
detection: A benchmark. In Computer Vision and Pattern
Recognition, 2009. CVPR 2009. IEEE Conference on , pages
304–311. IEEE, 2009.
[10] P. Dollar, C. Wojek, B. Schiele, and P. Perona. Pedestrian ´
detection: An evaluation of the state of the art. PAMI , 34,
2012.
[11] D. H. Eberly. 3D game engine design: a practical approach
to real-time computer graphics . CRC Press, 2006.
[12] M. Enzweiler and D. M. Gavrila. Monocular pedestrian detection: Survey and experiments. IEEE transactions on pattern analysis and machine intelligence , 31(12):2179–2195,
2009.
[13] A. Ess, B. Leibe, K. Schindler, , and L. van Gool. A mobile vision system for robust multi-person tracking. In IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR’08) . IEEE Press, June 2008.
[14] P. Fischer, A. Dosovitskiy, E. Ilg, P. Hausser, C. Hazırbas¸, ¨
V. Golkov, P. van der Smagt, D. Cremers, and T. Brox.
Flownet: Learning optical flow with convolutional networks.
arXiv preprint arXiv:1504.06852 , 2015.
[15] Y. Ganin and V. Lempitsky. Unsupervised domain adaptation
by backpropagation. arXiv preprint arXiv:1409.7495 , 2014.
[16] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle,
F. Laviolette, M. Marchand, and V. Lempitsky. Domainadversarial training of neural networks. Journal of Machine
Learning Research , 17(59):1–35, 2016.
[17] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu,
D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in Neural Information
Processing Systems , pages 2672–2680, 2014.
[18] K. Grauman, G. Shakhnarovich, and T. Darrell. Inferring
3d structure with a statistical image-based shape model. In
Computer Vision, 2003. Proceedings. Ninth IEEE International Conference on , pages 641–647. IEEE, 2003.
[19] H. Hattori, V. Naresh Boddeti, K. M. Kitani, and T. Kanade.
Learning scene-specific pedestrian detectors without real
data. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition , pages 3819–3827, 2015.
[20] M. Hejrati and D. Ramanan. Analysis by synthesis: 3d object recognition by object reconstruction. In 2014 IEEE Conference on Computer Vision and Pattern Recognition , pages
2449–2456. IEEE, 2014.
[21] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Imageto-image translation with conditional adversarial networks.
arXiv preprint arXiv:1611.07004 , 2016.
[22] D. P. Kingma and M. Welling. Auto-encoding variational
bayes. arXiv preprint arXiv:1312.6114 , 2013.
[23] K. Lai, L. Bo, and D. Fox. Unsupervised feature learning for
3d scene labeling. In 2014 IEEE International Conference on
Robotics and Automation (ICRA) , pages 3050–3057. IEEE,
2014.
[24] A. Lerer, S. Gross, and R. Fergus. Learning physical
intuition of block towers by example. arXiv preprint
arXiv:1603.01312 , 2016.
[25] R. Lienhart and J. Maydt. An extended set of haar-like features for rapid object detection. In Image Processing. 2002.
Proceedings. 2002 International Conference on , volume 1,
pages I–900. IEEE, 2002.
[26] M.-Y. Liu and O. Tuzel. Coupled generative adversarial networks. In Advances in Neural Information Processing Systems , pages 469–477, 2016.
[27] M. Long, Y. Cao, J. Wang, and M. I. Jordan. Learning transferable features with deep adaptation networks. In ICML ,
pages 97–105, 2015.
[28] M. M. Loper and M. J. Black. Opendr: An approximate differentiable renderer. In European Conference on Computer
Vision , pages 154–169. Springer, 2014.
[29] J. Marin, D. Vazquez, D. Ger ´ onimo, and A. M. L ´ opez. ´
Learning appearance in virtual scenarios for pedestrian
detection. In Computer Vision and Pattern Recognition
(CVPR), 2010 IEEE Conference on , pages 137–144. IEEE,
2010.
[30] N. Mayer, E. Ilg, P. Hausser, P. Fischer, D. Cremers, ¨
A. Dosovitskiy, and T. Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow
estimation. arXiv preprint arXiv:1512.02134 , 2015.
[31] M. Mirza and S. Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784 , 2014.
[32] Y. Movshovitz-Attias, Y. Sheikh, V. N. Boddeti, and Z. Wei.
3d pose-by-detection of vehicles via discriminatively reduced ensembles of correlation filters. In BMVC , 2014.
[33] W. Nam, P. Dollar, and J. H. Han. Local decorrelation for ´
improved pedestrian detection. In Advances in Neural Information Processing Systems , pages 424–432, 2014.
[34] B. Pepik, M. Stark, P. Gehler, and B. Schiele. Teaching 3d
geometry to deformable part models. In Computer Vision
and Pattern Recognition (CVPR), 2012 IEEE Conference on ,
pages 3362–3369. IEEE, 2012.
[35] L. Pishchulin, A. Jain, C. Wojek, M. Andriluka,
T. Thormahlen, and B. Schiele. Learning people detec- ¨
tion models from few training samples. In Computer Vision
and Pattern Recognition (CVPR), 2011 IEEE Conference on ,
pages 1473–1480. IEEE, 2011.
[36] M. Potamias and V. Athitsos. Nearest neighbor search methods for handshape recognition. In Proceedings of the 1st
international conference on PErvasive Technologies Related
to Assistive Environments , page 30. ACM, 2008.
[37] A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 , 2015.
[38] S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards
real-time object detection with region proposal networks. In
Advances in neural information processing systems , pages
91–99, 2015.
[39] D. J. Rezende, S. Mohamed, and D. Wierstra. Stochastic
backpropagation and approximate inference in deep generative models. arXiv preprint arXiv:1401.4082 , 2014.
[40] S. R. Richter, V. Vineet, S. Roth, and V. Koltun. Playing
for data: Ground truth from computer games. arXiv preprint
arXiv:1608.02192 , 2016.
[41] J. Romero, H. Kjellstrom, and D. Kragic. Hands in action: ¨
real-time 3d reconstruction of hands in interaction with objects. In Robotics and Automation (ICRA), 2010 IEEE International Conference on , pages 458–463. IEEE, 2010.
[42] G. Ros, L. Sellart, J. Materzynska, D. Vazquez, and A. M.
Lopez. The synthia dataset: A large collection of synthetic
images for semantic segmentation of urban scenes. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition , pages 3234–3243, 2016.
[43] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,
S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,
et al. Imagenet large scale visual recognition challenge.
International Journal of Computer Vision , 115(3):211–252,
2015.
[44] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen. Improved techniques for training gans.
arXiv preprint arXiv:1606.03498 , 2016.
[45] S. Satkin, J. Lin, and M. Hebert. Data-driven scene understanding from 3d models. 2012.
[46] B. Sun and K. Saenko. From virtual to reality: Fast adaptation of virtual object detectors to real domains. In BMVC ,
volume 1, page 3, 2014.
[47] P. Viola and M. Jones. Rapid object detection using a boosted
cascade of simple features. In Computer Vision and Pattern
Recognition, 2001. CVPR 2001. Proceedings of the 2001
IEEE Computer Society Conference on , volume 1, pages I–
511. IEEE, 2001.
[48] C. Wojek, S. Walk, and B. Schiele. Multi-cue onboard pedestrian detection. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on , pages 794–
801. IEEE, 2009.
[49] L. Zhang, L. Lin, X. Liang, and K. He. Is faster r-cnn doing
well for pedestrian detection? In European Conference on
Computer Vision , pages 443–457. Springer, 2016.
[50] Q. Zhu, M.-C. Yeh, K.-T. Cheng, and S. Avidan. Fast human
detection using a cascade of histograms of oriented gradients. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06) , volume 2,
pages 1491–1498. IEEE, 2006.