CVPR 2018 Papers

Our ability to train end-to-end systems for 3D human pose estimation from single images is currently constrained by the limited availability of 3D annotations for natural images. Most datasets are captured using Motion Capture (MoCap) systems in a studio setting and it is difficult to reach the variability of 2D human pose datasets, like MPII or LSP. To alleviate the need for accurate 3D ground truth, we propose to use a weaker supervision signal provided by the ordinal depths of human joints. This information can be acquired by human annotators for a wide range of images and poses. We showcase the effectiveness and flexibility of training Convolutional Networks (ConvNets) with these ordinal relations in different settings, always achieving competitive performance with ConvNets trained with accurate 3D joint coordinates. Additionally, to demonstrate the potential of the approach, we augment the popular LSP and MPII datasets with ordinal depth annotations. This extension allows us to present quantitative and qualitative evaluation in non-studio conditions. Simultaneously, these ordinal annotations can be easily incorporated in the training procedure of typical ConvNets for 3D human pose. Through this inclusion we achieve new state-of-the-art performance for the relevant benchmarks and validate the effectiveness of ordinal depth supervision for 3D human pose.

High-performance object detection relies on expensive convolutional networks to compute features, often leading to significant challenges in applications, e.g. those that re- quire detecting objects from video streams in real time. The key to this problem is to trade accuracy for efficiency in an effective way, i.e. reducing the computing cost while maintaining competitive performance. To seek a good balance, previous efforts usually focus on optimizing the model architectures. This paper explores an alternative approach, that is, to reallocate the computation over a scale-time space. The basic idea is to perform expensive detection sparsely and propagate the results across both scales and time with substantially cheaper networks, by exploiting the strong correlations among them. Specifically, we present a unified framework that integrates detection, temporal propagation, and across-scale refinement on a Scale-Time Lattice. On this framework, one can explore various strategies to balance performance and cost. Taking advantage of this flexibility, we further develop an adaptive scheme with the detector invoked on demand and thus obtain improved tradeoff. On ImageNet VID dataset, the proposed method can achieve a competitive mAP 79.6% at 20 fps, or 79.0% at 62 fps as a performance/speed tradeoff.

Recognizing facial action units (AUs) during spontaneous facial displays is a challenging problem. Most recently, Convolutional Neural Networks (CNNs) have shown promise for facial AU recognition, where predefined and fixed convolution filter sizes are employed. In order to achieve the best performance, the optimal filter size is often empirically found by conducting extensive experimental validation. Such a training process suffers from expensive training cost, especially as the network becomes deeper. This paper proposes a novel Optimized Filter Size CNN (OFS-CNN), where the filter sizes and weights of all convolutional layers are learned simultaneously from the training data along with learning convolution filters. Specifically, the filter size is defined as a continuous variable, which is optimized by minimizing the training loss. Experimental results on two AU-coded spontaneous databases have shown that the proposed OFS-CNN is capable of estimating optimal filter size for varying image resolution and outperforms traditional CNNs with the best filter size obtained by exhaustive search. The OFS-CNN also beats the CNN using multiple filter sizes and more importantly, is much more efficient during testing with the proposed forward-backward propagation algorithm.

We consider the problem of automatically generating sequences of structured-light patterns for active stereo triangulation of a static scene. Unlike existing approaches that use predetermined patterns and reconstruction algorithms tied to them, we generate patterns on the fly in response to generic specifications: number of patterns, projector-camera arrangement, workspace constraints, spatial frequency content, etc. Our pattern sequences are specifically optimized to minimize the expected rate of correspondence errors under those specifications for an unknown scene, and are coupled to a sequence-independent algorithm for per-pixel disparity estimation. To achieve this, we derive an objective function that is easy to optimize and follows from first principles within a maximum-likelihood framework. By minimizing it, we demonstrate automatic discovery of pattern sequences, in under three minutes on a laptop, that can outperform state-of-the-art triangulation techniques.

Motion representation plays a vital role in human action recognition in videos. In this study, we introduce a novel compact motion representation for video action recognition, named Optical Flow guided Feature (OFF), which enables the network to distill temporal information through a fast and robust approach. The OFF is derived from the definition of optical flow and is orthogonal to the optical flow. The derivation also provides theoretical support for using the difference between two frames. By directly calculating pixel-wise spatio-temporal gradients of the deep feature maps, the OFF could be embedded in any existing CNN based video action recognition framework with only a slight additional cost. It enables the CNN to extract spatio-temporal information, especially the temporal information between frames simultaneously. This simple but powerful idea is validated by experimental results. The network with OFF fed only by RGB inputs achieves a competitive accuracy of 93.3% on UCF-101, which is comparable with the result obtained by two streams (RGB and optical flow), but is 15 times faster in speed. Experimental results also show that OFF is complementary to other motion modalities such as optical flow. When the proposed method is plugged into the state-of-the-art video action recognition framework, it has 96.0% and 74.2% accuracy on UCF-101 and HMDB-51 respectively. The code for this project is available at: https://github.com/kevin-ssy/Optical-Flow-Guided-Feature

Learning based temporal action localization methods require vast amounts of training data. However, such large-scale video datasets, which are expected to capture the dynamics of every action category, are not only very expensive to acquire but are also not practical simply because there exists an uncountable number of action classes. This poses a critical restriction to the current methods when the training samples are few and rare (e.g. when the target action classes are not present in the current publicly available datasets). To address this challenge, we conceptualize a new example-based action detection problem where only a few examples are provided, and the goal is to find the occurrences of these examples in an untrimmed video sequence. Towards this objective, we introduce a novel one-shot action localization method that alleviates the need for large amounts of training samples. Our solution adopts the one-shot learning technique of Matching Network and utilizes correlations to mine and localize actions of previously unseen classes. We evaluate our one-shot action localization method on the THUMOS14 and ActivityNet datasets, of which we modified the configuration to fit our one-shot problem setup.

Deep Neural Networks (DNNs) have been demonstrated to perform exceptionally well on most recognition tasks such as image classification and segmentation. However, they have also been shown to be vulnerable to adversarial examples. This phenomenon has recently attracted a lot of attention but it has not been extensively studied on multiple, large-scale datasets and complex tasks such as semantic segmentation which often require more specialised networks with additional components such as CRFs, dilated convolutions, skip-connections and multiscale processing. In this paper, we present what to our knowledge is the first rigorous evaluation of adversarial attacks on modern semantic segmentation models, using two large-scale datasets. We analyse the effect of different network architectures, model capacity and multiscale processing, and show that many observations made on the task of classification do not always transfer to this more complex task. Furthermore, we show how mean-field inference in deep structured models and multiscale processing naturally implement recently proposed adversarial defenses. Our observations will aid future efforts in understanding and defending against adversarial examples. Moreover, in the shorter term, we show which segmentation models should currently be preferred in safety-critical applications due to their inherent robustness.

Convolutional networks (ConvNets) have become the dominant approach to semantic image segmentation. Producing accurate, pixel--level labels required for this task is a tedious and time consuming process; however, producing approximate, coarse labels could take only a fraction of the time and effort. We investigate the relationship between the quality of labels and the performance of ConvNets for semantic segmentation. We create a very large synthetic dataset with perfectly labeled street view scenes. From these perfect labels, we synthetically coarsen labels with different qualities and estimate human--hours required for producing them. We perform a series of experiments by training ConvNets with a varying number of training images and label quality. We found that the performance of ConvNets mostly depends on the time spent creating the training labels. That is, a larger coarsely--annotated dataset can yield the same performance as a smaller finely--annotated one. Furthermore, fine--tuning coarsely pre--trained ConvNets with few finely-annotated labels can yield comparable or superior performance to training it with a large amount of finely-annotated labels alone, at a fraction of the labeling cost. We demonstrate that our result is also valid for different network architectures, and various object classes in an urban scene.

Image dehazing deals with the removal of undesired loss of visibility in outdoor images due to the presence of fog. Retinex is a color vision model mimicking the ability of the Human Visual System to robustly discount varying illuminations when observing a scene under different spectral lighting conditions. Retinex has been widely explored in the computer vision literature for image enhancement and other related tasks. While these two problems are apparently unrelated, the goal of this work is to show that they can be connected by a simple linear relationship. Specifically, most Retinex-based algorithms have the characteristic feature of always increasing image brightness, which turns them into ideal candidates for effective image dehazing by directly applying Retinex to a hazy image whose intensities have been inverted. In this paper, we give theoretical proof that Retinex on inverted intensities is a solution to the image dehazing problem. Comprehensive qualitative and quantitative results indicate that several classical and modern implementations of Retinex can be transformed into competing image dehazing algorithms performing on pair with more complex fog removal methods, and can overcome some of the main challenges associated with this problem.

Many problems in image/video processing and computer vision require the computation of a dense k-nearest neighbor field (k-NNF) between two images. For each patch in a query image, the k-NNF determines the positions of the k most similar patches in a database image. With the introduction of the PatchMatch algorithm, Barnes et al. demonstrated that this large search problem can be approximated efficiently by collaborative search methods that exploit the local coherency of image patches. After its introduction, several variants of the original PatchMatch algorithm have been proposed, some of them reducing the computational time by two orders of magnitude. In this work we propose a theoretical framework for the analysis of PatchMatch and its variants, and apply it to derive bounds on their covergence rate. We consider a generic PatchMatch algorithm from which most specific instances found in the literature can be derived as particular cases. We also derive more specific bounds for two of these particular cases: the original PatchMatch and Coherency Sensitive Hashing. The proposed bounds are validated by contrasting them to the convergence observed in practice.

A vast majority of contemporary cameras employ rolling shutter (RS) mechanism to capture images. Due to the sequential mechanism, images acquired with a moving camera are subjected to rolling shutter effect which manifests as geometric distortions. In this work, we consider the specific scenario of a fast moving camera wherein the rolling shutter distortions not only are predominant but also become depth-dependent which in turn results in intra-frame occlusions. To this end, we develop a first-of-its-kind pipeline to recover the latent image of a 3D scene from a set of such RS distorted images. The proposed approach sequentially recovers both the camera motion and scene structure while accounting for RS and occlusion effects. Subsequently, we perform depth and occlusion-aware rectification of RS images to yield the desired latent image. Our experiments on synthetic and real image sequences reveal that the proposed approach achieves state-of-the-art results.

It has been recently shown that a convolutional neural network can learn optical flow estimation with unsuper- vised learning. However, the performance of the unsuper- vised methods still has a relatively large gap compared to its supervised counterpart. Occlusion and large motion are some of the major factors that limit the current unsuper- vised learning of optical flow methods. In this work we introduce a new method which models occlusion explicitly and a new warping way that facilitates the learning of large motion. Our method shows promising results on Flying Chairs, MPI-Sintel and KITTI benchmark datasets. Espe- cially on KITTI dataset where abundant unlabeled samples exist, our unsupervised method outperforms its counterpart trained with supervised learning.

Pedestrian detection has progressed significantly in the last years. However, occluded people are notoriously hard to detect, as their appearance varies substantially depending on a wide range of partial occlusions. In this paper, we aim to propose a simple and compact method based on the FasterRCNN architecture for occluded pedestrian detection. We start with interpreting CNN channel features of a pedestrian detector, and we find that different channels activate responses for different body parts respectively. These findings strongly motivate us to employ an attention mechanism across channels to represent various occlusion patterns in one single model, as each occlusion pattern can be formulated as some specific combination of body parts. Therefore, an attention network with self or external guidances is proposed as an add-on to the baseline FasterRCNN detector. When evaluating on the heavy occlusion subset, we achieve a significant improvement of 8pp to the baseline FasterRCNN detector on CityPersons and on Caltech we outperform the state-of-the-art method by 4pp.

We present a semantic part detection approach that effectively leverages object information. We use the object appearance and its class as indicators of what parts to expect. We also model the expected relative location of parts inside the objects based on their appearance. We achieve this with a new network module, called OffsetNet, that efficiently predicts a variable number of part locations within a given object. Our model incorporates all these cues to detect parts in the context of their objects. This leads to considerably higher performance for the challenging task of part detection compared to using part appearance alone (+5 mAP on the PASCAL-Part dataset). We also compare to other part detection methods on both PASCAL-Part and CUB200-2011 datasets.

We investigate the problem of object referring (OR) i.e. to localize a target object in a visual scene coming with a language description. Humans perceive the world more as continued video snippets than as static images, and describe objects not only by their appearance, but also by their spatio-temporal context and motion features. Humans also gaze at the object when they issue a referring expression. Existing works for OR mostly focus on static images only, which fall short in providing many such cues. This paper addresses OR in videos with language and human gaze. To that end, we present a new video dataset for OR, with 30,000 objects over 5,000 stereo video sequences annotated for their descriptions and gaze. We further propose a novel network model for OR in videos, by integrating appearance, motion, gaze, and spatio-temporal context into one network. Experimental results show that our method effectively utilizes motion cues, human gaze, and spatio-temporal context. Our method outperforms previous OR methods.

Deep neural networks trained using a softmax layer at the top and the cross-entropy loss are ubiquitous tools for image classification. Yet, this does not naturally enforce intra-class similarity nor inter-class margin of the learned deep representations. To simultaneously achieve these two goals, different solutions have been proposed in the literature, such as the pairwise or triplet losses. However, these carry the extra task of selecting pairs or triplets, and the extra computational burden of computing and learning for many combinations of them. In this paper, we propose a plug-and-play loss term for deep networks that explicitly reduces intra-class variance and enforces inter-class margin simultaneously, in a simple and elegant geometric manner. For each class, the deep features are collapsed into a learned linear subspace, or union of them, and inter-class subspaces are pushed to be as orthogonal as possible. Our proposed Orthogonal Low-rank Embedding (OLE) does not require carefully crafting pairs or triplets of samples for training, and works standalone as a classification loss, being the first reported deep metric learning framework of its kind. Because of the improved margin between features of different classes, the resulting deep networks generalize better, are more discriminative, and more robust. We demonstrate improved classification performance in general object recognition, plugging the proposed loss term into existing off-the-shelf architectures. In particular, we show the advantage of the proposed loss in the small data/model scenario, and we significantly advance the state-of-the-art on the Stanford STL-10 benchmark.

We present a novel approach to template matching that is efficient, can handle partial occlusions, and comes with provable performance guarantees. A key component of the method is a reduction that transforms the problem of searching a nearest neighbor among $N$ high-dimensional vectors, to searching neighbors among two sets of order $sqrt{N}$ vectors, which can be found efficiently using range search techniques. This allows for a quadratic improvement in search complexity, and makes the method scalable in handling large search spaces. The second contribution is a hashing scheme based on consensus set maximization, which allows us to handle occlusions. The resulting scheme can be seen as a randomized hypothesize-and-test algorithm, which is equipped with guarantees regarding the number of iterations required for obtaining an optimal solution with high probability. The predicted matching rates are validated empirically and the algorithm shows a significant improvement over the state-of-the-art in both speed and robustness to occlusions.

We are interested in enabling automatic 4D cinema by parsing physical and special effects from untrimmed movies. These include effects such as physical interactions, water splashing, light, and shaking, and are grounded to either a character in the scene or the camera. We collect a new dataset referred to as the Movie4D dataset which annotates over 9K effects in 63 movies. We propose a Conditional Random Field model atop a neural network that brings together visual and audio information, as well as semantics in the form of person tracks. Our model further exploits correlations of effects between different characters in the clip as well as across movie threads. We propose effect detection and classification as two tasks, and present results along with ablation studies on our dataset, paving the way towards 4D cinema in everyone’s homes.

Most recent semantic segmentation methods train deep convolutional neural networks with fully annotated masks requiring pixel-accuracy for good quality training. Common weakly-supervised approaches generate full masks from partial input (e.g. scribbles or seeds) using standard interactive segmentation methods as preprocessing. But, errors in such masks result in poorer training since standard loss functions (e.g. cross-entropy) do not distinguish seeds from potentially mislabeled other pixels. Inspired by the general ideas in semi-supervised learning, we address these problems via a new principled loss function evaluating network output with criteria standard in ``shallow'' segmentation, e.g. normalized cut. Unlike prior work, the cross entropy part of our loss evaluates only seeds where labels are known while normalized cut softly evaluates consistency of all pixels. We focus on normalized cut loss where dense Gaussian kernel is efficiently implemented in linear time by fast Bilateral filtering. Our normalized cut loss approach to segmentation brings the quality of weakly-supervised training significantly closer to fully supervised methods.