CVPR 2019 Papers

Temporal segmentation of long videos is an important problem, that has largely been tackled through supervised learning, often requiring large amounts of annotated training data. In this paper, we tackle the problem of self-supervised temporal segmentation that alleviates the need for any supervision in the form of labels (full supervision) or temporal ordering (weak supervision). We introduce a self-supervised, predictive learning framework that draws inspiration from cognitive psychology to segment long, visually complex videos into constituent events. Learning involves only a single pass through the training data. We also introduce a new adaptive learning paradigm that helps reduce the effect of catastrophic forgetting in recurrent neural networks. Extensive experiments on three publicly available datasets - Breakfast Actions, 50 Salads, and INRIA Instructional Videos datasets show the efficacy of the proposed approach. We show that the proposed approach outperforms weakly-supervised and unsupervised baselines by up to 24% and achieves competitive segmentation results compared to fully supervised baselines with only a single pass through the training data. Finally, we show that the proposed self-supervised learning paradigm learns highly discriminating features to improve action recognition.

In this paper, we address the problem of inferring the layout of complex road scenes given a single camera as input. To achieve that, we first propose a novel parameterized model of road layouts in a top-view representation, which is not only intuitive for human visualization but also provides an interpretable interface for higher-level decision making. Moreover, the design of our top-view scene model allows for efficient sampling and thus generation of large-scale simulated data, which we leverage to train a deep neural network to infer our scene model's parameters. Specifically, our proposed training procedure uses supervised domain-adaptation techniques to incorporate both simulated as well as manually annotated data. Finally, we design a Conditional Random Field (CRF) that enforces coherent predictions for a single frame and encourages temporal smoothness among video frames. Experiments on two public data sets show that: (1) Our parametric top-view model is representative enough to describe complex road scenes; (2) The proposed method outperforms baselines trained on manually-annotated or simulated data only, thus getting the best of both; (3) Our CRF is able to generate temporally smoothed while semantically meaningful results.

Neuroscience theory posits that the brain's visual system coarsely identifies broad object categories via neural activation patterns, with similar objects producing similar neural responses. Artificial neural networks also have internal activation behavior in response to stimuli. We hypothesize that networks exhibiting brain-like activation behavior will demonstrate brain-like characteristics, e.g., stronger generalization capabilities. In this paper we introduce a human-model similarity (HMS) metric, which quantifies the similarity of human fMRI and network activation behavior. To calculate HMS, representational dissimilarity matrices (RDMs) are created as abstractions of activation behavior, measured by the correlations of activations to stimulus pairs. HMS is then the correlation between the fMRI RDM and the neural network RDM across all stimulus pairs. We test the metric on unsupervised predictive coding networks, which specifically model visual perception, and assess the metric for statistical significance over a large range of hyperparameters. Our experiments show that networks with increased human-model similarity are correlated with better performance on two computer vision tasks: next frame prediction and object matching accuracy. Further, HMS identifies networks with high performance on both tasks. An unexpected secondary finding is that the metric can be employed during training as an early-stopping mechanism.

We propose novel neural temporal models for predicting and synthesizing human motion, achieving state-of-the-art in modeling long-term motion trajectories while being competitive with prior work in short-term prediction and requiring significantly less computation. Key aspects of our proposed system include: 1) a novel, two-level processing architecture that aids in generating planned trajectories, 2) a simple set of easily computable features that integrate derivative information, and 3) a novel multi-objective loss function that helps the model to slowly progress from simple next-step prediction to the harder task of multi-step, closed-loop prediction. Our results demonstrate that these innovations improve the modeling of long-term motion trajectories. Finally, we propose a novel metric, called Normalized Power Spectrum Similarity (NPSS), to evaluate the long-term predictive ability of motion synthesis models, complementing the popular mean-squared error (MSE) measure of Euler joint angles over time. We conduct a user study to determine if the proposed NPSS correlates with human evaluation of long-term motion more strongly than MSE and find that it indeed does. We release code and additional results (visualizations) for this paper at: https://github.com/cr7anand/neural_temporal_models

This paper proposes a new neural network based on SPD manifold learning for skeleton-based hand gesture recognition. Given the stream of hand's joint positions, our approach combines two aggregation processes on respectively spatial and temporal domains. The pipeline of our network architecture consists in three main stages. The first stage is based on a convolutional layer to increase the discriminative power of learned features. The second stage relies on different architectures for spatial and temporal Gaussian aggregation of joint features. The third stage learns a final SPD matrix from skeletal data. A new type of layer is proposed for the third stage, based on a variant of stochastic gradient descent on Stiefel manifolds. The proposed network is validated on two challenging datasets and shows state-of-the-art accuracies on both datasets.

Though deep learning techniques have made great progress in salient object detection recently, the predicted saliency maps still suffer from incomplete predictions due to the internal complexity of objects and inaccurate boundaries caused by strides in convolution and pooling operations. To alleviate these issues, we propose to train saliency detection networks by exploiting the supervision from not only salient object detection, but also foreground contour detection and edge detection. First, we leverage salient object detection and foreground contour detection tasks in an intertwined manner to generate saliency maps with uniform highlight. Second, the foreground contour and edge detection tasks guide each other simultaneously, thereby leading to preciser foreground contour prediction and reducing the local noises for edge prediction. In addition, we develop a novel mutual learning module (MLM) which serves as the building block of our method. Each MLM consists of multiple network branches trained in a mutual learning manner, which improves the performance by a large margin. Extensive experiments on seven challenging datasets demonstrate that the proposed method has delivered state-of-the-art results in both salient object detection and edge detection.

To reduce memory footprint and run-time latency, techniques such as neural net-work pruning and binarization have been explored separately. However, it is un-clear how to combine the best of the two worlds to get extremely small and efficient models. In this paper, we, for the first time, define the filter-level pruning problem for binary neural networks, which cannot be solved by simply migrating existing structural pruning methods for full-precision models. A novel learning-based approach is proposed to prune filters in our main/subsidiary network frame-work, where the main network is responsible for learning representative features to optimize the prediction performance, and the subsidiary component works as a filter selector on the main network. To avoid gradient mismatch when training the subsidiary component, we propose a layer-wise and bottom-up scheme. We also provide the theoretical and experimental comparison between our learning-based and greedy rule-based methods. Finally, we empirically demonstrate the effectiveness of our approach applied on several binary models, including binarizedNIN, VGG-11, and ResNet-18, on various image classification datasets. For bi-nary ResNet-18 on ImageNet, we use 78.6% filters but can achieve slightly better test error 49.87% (50.02%-0.15%) than the original model

The Convolutional Sparse Coding (CSC) model has recently gained considerable traction in the signal and image processing communities. By providing a global, yet tractable, model that operates on the whole image, the CSC was shown to overcome several limitations of the patch-based sparse model while achieving superior performance in various applications. Contemporary methods for pursuit and learning the CSC dictionary often rely on the Alternating Direction Method of Multipliers (ADMM) in the Fourier domain for the computational convenience of convolutions, while ignoring the local characterizations of the image. In this work we propose a new and simple approach that adopts a localized strategy, based on the Block Coordinate Descent algorithm. The proposed method, termed Local Block Coordinate Descent (LoBCoD), operates locally on image patches. Furthermore, we introduce a novel stochastic gradient descent version of LoBCoD for training the convolutional filters. This Stochastic-LoBCoD leverages the benefits of online learning, while being applicable even to a single training image. We demonstrate the advantages of the proposed algorithms for image inpainting and multi-focus image fusion, achieving state-of-the-art results.

This paper studies the structure of a deep convolutional neural network to predict the foreground alpha matte by taking a single RGB image as input. Our network is fully convolutional with two decoder branches for the foreground and background classification respectively. Then a fusion branch is used to integrate the two classification results which gives rise to alpha values as the soft segmentation result. This design provides more degrees of freedom than a single decoder branch for the network to obtain better alpha values during training. The network can implicitly produce trimaps without user interaction, which is easy to use for novices without expertise in digital matting. Experimental results demonstrate that our network can achieve high-quality alpha mattes for various types of objects and outperform the state-of-the-art CNN-based image matting methods on the human image matting task.

The linear and non-flexible nature of deep convolutional models makes them vulnerable to carefully crafted adversarial perturbations. To tackle this problem, we propose a non-linear radial basis convolutional feature mapping by learning a Mahalanobis-like distance function. Our method then maps the convolutional features onto a linearly well-separated manifold, which prevents small adversarial perturbations from forcing a sample to cross the decision boundary. We test the proposed method on three publicly available image classification and segmentation datasets namely, MNIST, ISBI ISIC 2017 skin lesion segmentation, and NIH Chest X-Ray-14. We evaluate the robustness of our method to different gradient (targeted and untargeted) and non-gradient based attacks and compare it to several non-gradient masking defense strategies. Our results demonstrate that the proposed method can increase the resilience of deep convolutional neural networks to adversarial perturbations without accuracy drop on clean data.

We propose a new flexible deep convolutional neural network (convnet) to perform fast neural style transfers. Our network is trained to solve approximately, but rapidly, the artistic style transfer problem of [Gatys et al.] for arbritary styles. While solutions already exist, our network is uniquely flexible by design: it can be manipulated at runtime to enforce new constraints on the final output. As examples, we show that it can be modified to perform tasks such as fast photorealistic style transfer, or fast video style transfer with short term consistency, with no retraining. This flexibility stems from the proposed architecture which is obtained by unrolling the gradient descent algorithm used in [Gatys et al.]. Regularisations added to [Gatys et al.] to solve a new task can be reported on-the-fly in our network, even after training.

The sparse generalized eigenvalue problem arises in a number of standard and modern statistical learning models, including sparse principal component analysis, sparse Fisher discriminant analysis, and sparse canonical correlation analysis. However, this problem is difficult to solve since it is NP-hard. In this paper, we consider a new effective decomposition method to tackle this problem. Specifically, we use random or/and swapping strategies to find a working set and perform global combinatorial search over the small subset of variables. We consider a bisection search method and a coordinate descent method for solving the quadratic fractional programming subproblem. In addition, we provide some theoretical analysis for the proposed method. Our experiments on synthetic data and real-world data have shown that our method significantly and consistently outperforms existing solutions in term of accuracy.

Face anti-spoofing is essential to prevent face recognition systems from a security breach. Much of the progresses have been made by the availability of face anti-spoofing benchmark datasets in recent years. However, existing face anti-spoofing benchmarks have limited number of subjects (<=170) and modalities (<=2), which hinder the further development of the academic community. To facilitate face anti-spoofing research, we introduce a large-scale multi-modal dataset, namely CASIA-SURF, which is the largest publicly available dataset for face anti-spoofing in terms of both subjects and visual modalities. Specifically, it consists of 1,000 subjects with 21,000 videos and each sample has 3 modalities (i.e., RGB, Depth and IR). We also provide a measurement set, evaluation protocol and training/validation/testing subsets, developing a new benchmark for face anti-spoofing. Moreover, we present a new multi-modal fusion method as baseline, which performs feature re-weighting to select the more informative channel features while suppressing the less useful ones for each modal. Extensive experiments have been conducted on the proposed dataset to verify its significance and generalization capability. The dataset is available at https://sites.google.com/qq.com/chalearnfacespoofingattackdete/.

In this paper, we present a method to utilize 2D-2D point matches between images taken during different image conditions to train a convolutional neural network for semantic segmentation. Enforcing label consistency across the matches makes the final segmentation algorithm robust to seasonal changes. We describe how these 2D-2D matches can be generated with little human interaction by geometrically matching points from 3D models built from images. Two cross-season correspondence datasets are created providing 2D-2D matches across seasonal changes as well as from day to night. The datasets are made publicly available to facilitate further research. We show that adding the correspondences as extra supervision during training improves the segmentation performance of the convolutional neural network, making it more robust to seasonal changes and weather conditions.

Style transfer has recently received a lot of attention, since it allows to study fundamental challenges in image understanding and synthesis. Recent work has significantly improved the representation of color and texture and com- putational speed and image resolution. The explicit transformation of image content has, however, been mostly neglected: while artistic style affects formal characteristics of an image, such as color, shape or texture, it also deforms, adds or removes content details. This paper explicitly focuses on a content-and style-aware stylization of a content image. Therefore, we introduce a content transformation module between the encoder and decoder. Moreover, we utilize similar content appearing in photographs and style samples to learn how style alters content details and we generalize this to other class details. Additionally, this work presents a novel normalization layer critical for high resolution image synthesis. The robustness and speed of our model enables a video stylization in real-time and high definition. We perform extensive qualitative and quantitative evaluations to demonstrate the validity of our approach.

Most of the existing work on automatic facial expression analysis focuses on discrete emotion recognition, or facial action unit detection. However, facial expressions do not always fall neatly into pre-defined semantic categories. Also, the similarity between expressions measured in the action unit space need not correspond to how humans perceive expression similarity. Different from previous work, our goal is to describe facial expressions in a continuous fashion using a compact embedding space that mimics human visual preferences. To achieve this goal, we collect a large-scale faces-in-the-wild dataset with human annotations in the form: Expressions A and B are visually more similar when compared to expression C, and use this dataset to train a neural network that produces a compact (16-dimensional) expression embedding. We experimentally demonstrate that the learned embedding can be successfully used for various applications such as expression retrieval, photo album summarization, and emotion recognition. We also show that the embedding learned using the proposed dataset performs better than several other embeddings learned using existing emotion or action unit datasets.

The deep image prior was recently introduced as a prior for natural images. It represents images as the output of a convolutional network with random inputs. For "inference", gradient descent is performed to adjust network parameters to make the output match observations. This approach yields good performance on a range of image reconstruction tasks. We show that the deep image prior is asymptotically equivalent to a stationary Gaussian process prior in the limit as the number of channels in each layer of the network goes to infinity, and derive the corresponding kernel. This informs a Bayesian approach to inference. We show that by conducting posterior inference using stochastic gradient Langevin dynamics we avoid the need for early stopping, which is a drawback of the current approach, and improve results for denoising and impainting tasks. We illustrate these intuitions on a number of 1D and 2D signal reconstruction tasks.

In many robotics and VR/AR applications, 3D-videos are readily-available input sources (a sequence of depth images, or LIDAR scans). However, in many cases, the 3D-videos are processed frame-by-frame either through 2D convnets or 3D perception algorithms. In this work, we propose 4-dimensional convolutional neural networks for spatio-temporal perception that can directly process such 3D-videos using high-dimensional convolutions. For this, we adopt sparse tensors and propose generalized sparse convolutions that encompass all discrete convolutions. To implement the generalized sparse convolution, we create an open-source auto-differentiation library for sparse tensors that provides extensive functions for high-dimensional convolutional neural networks. We create 4D spatio-temporal convolutional neural networks using the library and validate them on various 3D semantic segmentation benchmarks and proposed 4D datasets for 3D-video perception. To overcome challenges in 4D space, we propose the hybrid kernel, a special case of the generalized sparse convolution, and trilateral-stationary conditional random fields that enforce spatio-temporal consistency in the 7D space-time-chroma space. Experimentally, we show that a convolutional neural network with only generalized 3D sparse convolutions can outperform 2D or 2D-3D hybrid methods by a large margin. Also, we show that on 3D-videos, 4D spatio-temporal convolutional neural networks are robust to noise and outperform the 3D convolutional neural network.

Applications in virtual and augmented reality create a demand for rapid creation and easy access to large sets of 3D models. An effective way to address this demand is to edit or deform existing 3D models based on a reference, e.g., a 2D image which is very easy to acquire. Given such a source 3D model and a target which can be a 2D image, 3D model, or a point cloud acquired as a depth scan, we introduce 3DN, an end-to-end network that deforms the source model to resemble the target. Our method infers per-vertex offset displacements while keeping the mesh connectivity of the source model fixed. We present a training strategy which uses a novel differentiable operation, mesh sampling operator, to generalize our method across source and target models with varying mesh densities. Mesh sampling operator can be seamlessly integrated into the network to handle meshes with different topologies. Qualitative and quantitative results show that our method generates higher quality results compared to the state-of-the art learning-based methods for 3D shape generation.