CVPR 2021 Papers

When in a new situation or geographical location, human drivers have an extraordinary ability to watch others and learn maneuvers that they themselves may have never performed. In contrast, existing techniques for learning to drive preclude such a possibility as they assume direct access to an instrumented ego-vehicle with fully known observations and expert driver actions. However, such measurements cannot be directly accessed for the non-ego vehicles when learning by watching others. Therefore, in an application where data is regarded as a highly valuable asset, current approaches completely discard the vast portion of the training data that can be potentially obtained through indirect observation of surrounding vehicles. Motivated by this key insight, we propose the Learning by Watching (LbW) framework which enables learning a driving policy without requiring full knowledge of neither the state nor expert actions. To increase its data, i.e., with new perspectives and maneuvers, LbW makes use of the demonstrations of other vehicles in a given scene by (1) transforming the ego-vehicle's observations to their points of view, and (2) inferring their expert actions. Our LbW agent learns more robust driving policies while enabling data-efficient learning, including quick adaptation of the policy to rare and novel scenarios. In particular, LbW drives robustly even with a fraction of available driving data required by existing methods, achieving an average success rate of 92% on the original CARLA benchmark with only 30 minutes of total driving data and 82% with only 10 minutes.

Recently, language-guided global image editing draws increasing attention with growing application potentials. However, previous GAN-based methods are not only confined to domain-specific, low-resolution data but also lacking in interpretability. To overcome the collective difficulties, we develop a text-to-operation model to map the vague editing language request into a series of editing operations, e.g., change contrast, brightness, and saturation. Each operation is interpretable and differentiable. Furthermore, the only supervision in the task is the target image, which is insufficient for a stable training of sequential decisions. Hence, we propose a novel operation planning algorithm to generate possible editing sequences from the target image as pseudo ground truth. Comparison experiments on the newly collected MA5k-Req dataset and GIER dataset show the advantages of our methods. Code is available at https://github.com/jshi31/T2ONet.

We present a self-supervised approach for learning video representations using temporal video alignment as a pretext task, while exploiting both frame-level and video-level information. We leverage a novel combination of temporal alignment loss and temporal regularization terms, which can be used as supervision signals for training an encoder network. Specifically, the temporal alignment loss (i.e., Soft-DTW) aims for the minimum cost for temporally aligning videos in the embedding space. However, optimizing solely for this term leads to trivial solutions, particularly, one where all frames get mapped to a small cluster in the embedding space. To overcome this problem, we propose a temporal regularization term (i.e., Contrastive-IDM) which encourages different frames to be mapped to different points in the embedding space. Extensive evaluations on various tasks, including action phase classification, action phase progression, and fine-grained frame retrieval, on three datasets, namely Pouring, Penn Action, and IKEA ASM, show superior performance of our approach over state-of-the-art methods for self-supervised representation learning from videos. In addition, our method provides significant performance gain where labeled data is lacking.

A general approach for handling hyperspectral image (HSI) denoising issue is to impose weights on different HSI pixels to suppress negative influence brought by noisy elements. Such weighting scheme, however, largely depends on the prior understanding or subjective distribution assumption on HSI noises, making them easily biased to complicated real noises, and hardly generalizable to diverse practical scenarios. Against this issue, this paper proposes a new scheme aiming to capture general weighting principle in a data-driven manner. Specifically, such weighting principle is delivered by an explicit function, called hyperweight-net (HWnet), mapping from an input noisy image to its properly imposed weights. A Bayesian framework, as well as a variational inference algorithm, for inferring HWnet parameters is elaborately designed, expecting to extract the latent weighting rule for general diverse and complicated noisy HSIs. Comprehensive experiments substantiate that the learned HWnet can be not only finely generalized to different noise types from those used in training, but also effectively transferred to other weighted models for the issue. Besides, as a sounder guidance, HWnet can help to more faithfully and robustly achieve deep hyperspectral prior(DHP). Specially, the extracted weights by HWnet are verified to be able to effectively capture complex noise knowledge underlying input HSI, revealing its working insight in experiments.

State-of-the-art subspace clustering methods are based on the self-expressive model, which represents each data point as a linear combination of other data points. However, such methods are designed for a finite sample dataset and lack the ability to generalize to out-of-sample data. Moreover, since the number of self-expressive coefficients grows quadratically with the number of data points, their ability to handle large-scale datasets is often limited. In this paper, we propose a novel framework for subspace clustering, termed Self-Expressive Network (SENet), which employs a properly designed neural network to learn a self-expressive representation of the data. We show that our SENet can not only learn the self-expressive coefficients with desired properties on the training data, but also handle out-of-sample data. Besides, we show that SENet can also be leveraged to perform subspace clustering on large-scale datasets. Extensive experiments conducted on synthetic data and real world benchmark data validate the effectiveness of the proposed method. In particular, SENet yields highly competitive performance on MNIST, Fashion MNIST and Extended MNIST and state-of-the-art performance on CIFAR-10.

The recent trend in multiple object tracking (MOT) is heading towards leveraging deep learning to boost the tracking performance. However, it is not trivial to solve the data-association problem in an end-to-end fashion. In this paper, we propose a novel proposal-based learnable framework, which models MOT as a proposal generation, proposal scoring and trajectory inference paradigm on an affinity graph. This framework is similar to the two-stage object detector Faster RCNN, and can solve the MOT problem in a data-driven way. For proposal generation, we propose an iterative graph clustering method to reduce the computational cost while maintaining the quality of the generated proposals. For proposal scoring, we deploy a trainable graph-convolutional-network (GCN) to learn the structural patterns of the generated proposals and rank them according to the estimated quality scores. For trajectory inference, a simple deoverlapping strategy is adopted to generate tracking output while complying with the constraints that no detection can be assigned to more than one track. We experimentally demonstrate that the proposed method achieves a clear performance improvement in both MOTA and IDF1 with respect to previous state-of-the-art on two public benchmarks. Our code is available at https://github.com/daip13/LPC_MOT.git.

Deblurring night blurry images is difficult, because the common-used blur model based on the linear convolution operation does not hold in this situation due to the influence of saturated pixels. In this paper, we propose a non-blind deblurring network (NBDN) to restore night blurry images. To mitigate the side effects brought by the pixels that violate the blur model, we develop a confidence estimation unit (CEU) to estimate a map which ensures smaller contributions of these pixels to the deconvolution steps that are further optimized by the conjugate gradient (CG) method. Moreover, unlike the existing methods using manually tuned hyper-parameters in their frameworks, we propose a hyper-parameter estimation unit (HPEU) to adaptively estimate hyper-parameters for better image restoration . The experimental results demonstrate that the proposed network performs favorably against state-of-the-art algorithms both quantitatively and qualitatively.

The facial expression analysis requires a compact and identity-ignored expression representation. In this paper, we model the expression as the deviation from the identity by a subtraction operation, extracting a continuous and identity-invariant expression embedding. We propose a Deviation Learning Network (DLN) with a pseudo-siamese structure to extract the deviation feature vector. To reduce the optimization difficulty caused by additional fully connection layers, DLN directly provides high-order polynomial to nonlinearly project the high-dimensional feature to a low-dimensional manifold. Taking label noise into account, we add a crowd layer to DLN for robust embedding extraction. Also, to achieve a more compact representation, we use hierarchical annotation for data augmentation. We evaluate our facial expression embedding on the FEC validation set. The quantitative results prove that we achieve the state-of-the-art, both in terms of fine-grained and identity-invariant property. We further conduct extensive experiments to show that our expression embedding is of high quality for emotion recognition, image retrieval, and face manipulation.

We introduce a novel representation learning method to disentangle pose-dependent as well as view-dependent factors from 2D human poses. The method trains a network using cross-view mutual information maximization (CV-MIM) which maximizes mutual information of the same pose performed from different viewpoints in a contrastive learning manner. We further propose two regularization terms to ensure disentanglement and smoothness of the learned representations. The resulting pose representations can be used for cross-view action recognition. To evaluate the power of the learned representations, in addition to the conventional fully-supervised action recognition settings, we introduce a novel task called single-shot cross-view action recognition. This task trains models with actions from only one single viewpoint while models are evaluated on poses captured from all possible viewpoints. We evaluate the learned representations on standard benchmarks for action recognition, and show that (i) CV-MIM performs competitively compared with the state-of-the-art models in the fully-supervised scenarios; (ii) CV-MIM outperforms other competing methods by a large margin in the single-shot cross-view setting; (iii) and the learned representations can significantly boost the performance when reducing the amount of supervised training data. Our code is made publicly available at https://github.com/google-research/google-research/tree/master/poem.

Efficient 3D space sampling to represent an underlying3D object/scene is essential for 3D vision, robotics, and be-yond. A standard approach is to explicitly sample a densecollection of views and formulate it as a view selection prob-lem, or, more generally, a set cover problem. In this paper,we introduce a novel approach that avoids dense view sam-pling. The key idea is to learn a view prediction networkand a trainable aggregation module that takes the predictedviews as input and outputs an approximation of their genericscores (e.g., surface coverage, viewing angle from surfacenormals). This methodology allows us to turn the set coverproblem (or multi-view representation optimization) into acontinuous optimization problem. We then explain how toeffectively solve the induced optimization problem using con-tinuation, i.e., aggregating a hierarchy of smoothed scoringmodules. Experimental results show that our approach ar-rives at similar or better solutions with about 10 x speed upin running time, comparing with the standard methods.

Humans possess a unique social cognition capability; nonverbal communication can convey rich social information among agents. In contrast, such crucial social characteristics are mostly missing in the existing scene understanding literature. In this paper, we incorporate different nonverbal communication cues (e.g., gaze, human poses, and gestures) to represent, model, learn, and infer agents' mental states from pure visual inputs. Crucially, such a mental representation takes the agent's belief into account so that it represents what the true world state is and infers the beliefs in each agent's mental state, which may differ from the true world states. By aggregating different beliefs and true world states, our model essentially forms "five minds" during the interactions between two agents. This "five minds" model differs from prior works that infer beliefs in an infinite recursion; instead, agents' beliefs are converged into a "common mind". Based on this representation, we further devise a hierarchical energy-based model that jointly tracks and predicts all five minds. From this new perspective, a social event is interpreted by a series of nonverbal communication and belief dynamics, which transcends the classic keyframe video summary. In the experiments, we demonstrate that using such a social account provides a better video summary on videos with rich social interactions compared with state-of-the-art keyframe video summary methods.

Since its inception in 2015, Style Transfer has focused on texturing a content image using an art exemplar. Recently, the geometric changes that artists make have been acknowledged as an important component of style. Our contribution is to propose a neural network that, uniquely, learns a mapping from a 4D array of inter-feature distances to a non-parametric 2D warp field. The system is generic in not being limited by semantic class, a single learned model will suffice; all examples in this paper are output from one model. Our approach combines the benefits of the high speed of Liu et al. with the non-parametric warping of Kim et al. Furthermore, our system extends the normal NST paradigm: although it can be used with a single exemplar, we also allow two style exemplars: one for texture and another for geometry. This supports far greater flexibility in use cases than single exemplars can provide.

Appearance-based detectors achieve remarkable performance on common scenes, benefiting from high-capacity models and massive annotated data, but tend to fail for scenarios that lack training data. Geometric motion segmentation algorithms, however, generalize to novel scenes, but have yet to achieve comparable performance to appearance-based ones, due to noisy motion estimations and degenerate motion configurations. To combine the best of both worlds, we propose a modular network, whose architecture is motivated by a geometric analysis of what independent object motions can be recovered from an ego-motion field. It takes two consecutive frames as input and predicts segmentation masks for the background and multiple rigidly moving objects, which are then parameterized by 3D rigid transformations. Our method achieves state-of-the-art performance for rigid motion segmentation on KITTI and Sintel. The inferred rigid motions lead to a significant improvement for depth and scene flow estimation.

We address the problem of unsupervised localization of key-steps and feature learning in instructional videos using both visual and language instructions. Our key observation is that the sequences of visual and linguistic key-steps are weakly aligned: there is an ordered one-to-one correspondence between most visual and language key-steps, while some key-steps in one modality are absent in the other. To recover the two sequences, we develop an ordered prototype learning module, which extracts visual and linguistic prototypes representing key-steps. On the other hand, to find weak alignment and perform feature learning, we develop a differentiable weak sequence alignment (DWSA) method that finds ordered one-to-one matching between sequences while allowing some items in a sequence to stay unmatched. We develop an efficient forward and backward algorithm for computing the alignment and the loss derivative with respect to parameters of visual and language feature learning modules. By experiments on two instructional video datasets, we show that our method significantly improves the state of the art.

Most of the existing deep learning-based dehazing methods are trained and evaluated on the image dehazing datasets, where the dehazed images are generated by only exploiting the information from the corresponding hazy ones. On the other hand, the video dehazing algorithms, which can acquire more satisfying dehazing results by exploiting the temporal redundancy from neighborhood hazy frames, receive less attention due to the absence of the video dehazing datasets. Therefore, we propose the first REal-world VIdeo DEhazing (REVIDE) dataset which can be used for the supervised learning of the video dehazing algorithms. By utilizing a well-designed video acquisition system, we can capture paired real-world hazy and haze-free videos that are perfectly aligned by recording the same scene (with or without haze) twice. Considering the challenge of exploiting temporal redundancy among the hazy frames, we also develop a Confidence Guided and Improved Deformable Network (CG-IDN) for video dehazing. The experiments demonstrate that the hazy scenes in the REVIDE dataset are more realistic than the synthetic datasets and the proposed algorithm also performs favorably against state-of-the-art dehazing methods.

We present an approach for encoding visual task relationships to improve model performance in an Unsupervised Domain Adaptation (UDA) setting. Semantic segmentation and monocular depth estimation are shown to be complementary tasks; in a multi-task learning setting, a proper encoding of their relationships can further improve performance on both tasks. Motivated by this observation, we propose a novel Cross-Task Relation Layer (CTRL), which encodes task dependencies between the semantic and depth predictions. To capture the cross-task relationships, we propose a neural network architecture that contains task-specific and cross-task refinement heads. Furthermore, we propose an Iterative Self-Learning (ISL) training scheme, which exploits semantic pseudo-labels to provide extra supervision on the target domain. We experimentally observe improvements in both tasks' performance because the complementary information present in these tasks is better captured. Specifically, we show that: (1) our approach improves performance on all tasks when they are complementary and mutually dependent; (2) the CTRL helps to improve both semantic segmentation and depth estimation tasks performance in the challenging UDA setting; (3) the proposed ISL training scheme further improves the semantic segmentation performance. The implementation is available at https://github.com/susaha/ctrl-uda.

Despite significant progress in monocular depth estimation in the wild, recent state-of-the-art methods cannot be used to recover accurate 3D scene shape due to an unknown depth shift induced by shift-invariant reconstruction losses used in mixed-data depth prediction training, and possible unknown camera focal length. We investigate this problem in detail and propose a two-stage framework that first predicts depth up to an unknown scale and shift from a single monocular image, and then use 3D point cloud encoders to predict the missing depth shift and focal length that allow us to recover a realistic 3D scene shape. In addition, we propose an image-level normalized regression loss and a normal-based geometry loss to enhance depth prediction models trained on mixed datasets. We test our depth model on nine unseen datasets and achieve state-of-the-art performance on zero-shot dataset generalization. Code is available at:https://git.io/Depth.

Event cameras are novel sensors that capture the dynamics of a scene asynchronously. Such cameras record event streams with much shorter response latency than images captured by conventional cameras, and are also highly sensitive to intensity change, which is brought by the triggering mechanism of events. On the basis of these two features, previous works attempt to reconstruct high speed and high dynamic range (HDR) videos from events. However, these works either suffer from unrealistic artifacts, or cannot provide sufficiently high frame rate. In this paper, we present a convolutional recurrent neural network which takes a sequence of neighboring events to reconstruct high speed HDR videos, and temporal consistency is well considered to facilitate the training process. In addition, we setup a prototype optical system to collect a real-world dataset with paired high speed HDR videos and event streams, which will be made publicly accessible for future researches in this field. Experimental results on both simulated and real scenes verify that our method can generate high speed HDR videos with high quality, and outperform the state-of-the-art reconstruction methods.

This paper proposes a framework for the interactive video object segmentation (VOS) in the wild where users can choose some frames for annotations iteratively. Then, based on the user annotations, a segmentation algorithm refines the masks. The previous interactive VOS paradigm selects the frame with some worst evaluation metric, and the ground truth is required for calculating the evaluation metric, which is impractical in the testing phase. In contrast, in this paper, we advocate that the frame with the worst evaluation metric may not be exactly the most valuable frame that leads to the most performance improvement across the video. Thus, we formulate the frame selection problem in the interactive VOS as a Markov Decision Process, where an agent is learned to recommend the frame under a deep reinforcement learning framework. The learned agent can automatically determine the most valuable frame, making the interactive setting more practical in the wild. Experimental results on the public datasets show the effectiveness of our learned agent without any changes to the underlying VOS algorithms. Our data, code, and models are available at https://github.com/svip-lab/IVOS-W.