CVPR 2022 Papers

Generating speech-consistent body and gesture movements is a long-standing problem in virtual avatar creation. Previous studies often synthesize pose movement in a holistic manner, where poses of all joints are generated simultaneously. Such a straightforward pipeline fails to generate fine-grained co-speech gestures. One observation is that the hierarchical semantics in speech and the hierarchical structures of human gestures can be naturally described into multiple granularities and associated together. To fully utilize the rich connections between speech audio and human gestures, we propose a novel framework named Hierarchical Audio-to-Gesture (HA2G) for co-speech gesture generation. In HA2G, a Hierarchical Audio Learner extracts audio representations across semantic granularities. A Hierarchical Pose Inferer subsequently renders the entire human pose gradually in a hierarchical manner. To enhance the quality of synthesized gestures, we develop a contrastive learning strategy based on audio-text alignment for better audio representations. Extensive experiments and human evaluation demonstrate that the proposed method renders realistic co-speech gestures and outperforms previous methods in a clear margin. Project page: https://alvinliu0.github.io/projects/HA2G.

We introduce a novel formulation for guided super-resolution. Its core is a differentiable optimisation layer that operates on a learned affinity graph. The learned graph potentials make it possible to leverage rich contextual information from the guide image, while the explicit graph optimisation within the architecture guarantees rigorous fidelity of the high-resolution target to the low-resolution source. With the decision to employ the source as a constraint rather than only as an input to the prediction, our method differs from state-of-the-art deep architectures for guided super-resolution, which produce targets that, when downsampled, will only approximately reproduce the source. This is not only theoretically appealing, but also produces crisper, more natural-looking images. A key property of our method is that, although the graph connectivity is restricted to the pixel lattice, the associated edge potentials are learned with a deep feature extractor and can encode rich context information over large receptive fields. By taking advantage of the sparse graph connectivity, it becomes possible to propagate gradients through the optimisation layer and learn the edge potentials from data. We extensively evaluate our method on several datasets, and consistently outperform recent baselines in terms of quantitative reconstruction errors, while also delivering visually sharper outputs. Moreover, we demonstrate that our method generalises particularly well to new datasets not seen during training.

Natural videos provide rich visual contents for self-supervised learning. Yet most existing approaches for learning spatio-temporal representations rely on manually trimmed videos, leading to limited diversity in visual patterns and limited performance gain. In this work, we aim to learn representations by leveraging more abundant information in untrimmed videos. To this end, we propose to learn a hierarchy of consistencies in videos, i.e., visual consistency and topical consistency, corresponding respectively to clip pairs that tend to be visually similar when separated by a short time span and share similar topics when separated by a long time span. Specifically, a hierarchical consistency learning framework HiCo is presented, where the visually consistent pairs are encouraged to have the same representation through contrastive learning, while the topically consistent pairs are coupled through a topical classifier that distinguishes whether they are topicrelated. Further, we impose a gradual sampling algorithm for proposed hierarchical consistency learning, and demonstrate its theoretical superiority. Empirically, we show that not only HiCo can generate stronger representations on untrimmed videos, it also improves the representation quality when applied to trimmed videos. This is in contrast to standard contrastive learning that fails to learn appropriate representations from untrimmed videos.

Semi-supervised video action recognition tends to enable deep neural networks to achieve remarkable performance even with very limited labeled data. However, existing methods are mainly transferred from current image-based methods (e.g., FixMatch). Without specifically utilizing the temporal dynamics and inherent multimodal attributes, their results could be suboptimal. To better leverage the encoded temporal information in videos, we introduce temporal gradient as an additional modality for more attentive feature extraction in this paper. To be specific, our method explicitly distills the fine-grained motion representations from temporal gradient (TG) and imposes consistency across different modalities (i.e., RGB and TG). The performance of semi-supervised action recognition is significantly improved without additional computation or parameters during inference. Our method achieves the state-of-the-art performance on three video action recognition benchmarks (i.e., Kinetics-400, UCF-101, and HMDB-51) under several typical semi-supervised settings (i.e., different ratios of labeled data).

Saliency detection with light field images is becoming attractive given the abundant cues available, however, this comes at the expense of large-scale pixel level annotated data which is expensive to generate. In this paper, we propose to learn light field saliency from pixel-level noisy labels obtained from unsupervised hand crafted featured-based saliency methods. Given this goal, a natural question is: can we efficiently incorporate the relationships among light field cues while identifying clean labels in a unified framework? We address this question by formulating the learning as a joint optimization of intra light field features fusion stream and inter scenes correlation stream to generate the predictions. Specially, we first introduce a pixel forgetting guided fusion module to mutually enhance the light field features and exploit pixel consistency across iterations to identify noisy pixels. Next, we introduce a cross scene noise penalty loss for better reflecting latent structures of training data and enabling the learning to be invariant to noise. Extensive experiments on multiple benchmark datasets demonstrate the superiority of our framework showing that it learns saliency prediction comparable to state-of-the-art fully supervised light field saliency methods. Our code is available at https://github.com/OLobbCode/NoiseLF.

In this paper, we present a system to train driving policies from experiences collected not just from the ego-vehicle, but all vehicles that it observes. This system uses the behaviors of other agents to create more diverse driving scenarios without collecting additional data. The main difficulty in learning from other vehicles is that there is no sensor information. We use a set of supervisory tasks to learn an intermediate representation that is invariant to the viewpoint of the controlling vehicle. This not only provides a richer signal at training time but also allows more complex reasoning during inference. Learning how all vehicles drive helps predict their behavior at test time and can avoid collisions. We evaluate this system in closed-loop driving simulations. Our system outperforms all prior methods on the public CARLA Leaderboard by a wide margin, improving driving score by 25 and route completion rate by 24 points.

Recently, fairness-aware learning have become increasingly crucial, but most of those methods operate by assuming the availability of fully annotated demographic group labels. We emphasize that such assumption is unrealistic for real-world applications since group label annotations are expensive and can conflict with privacy issues. In this paper, we consider a more practical scenario, dubbed as Algorithmic Group Fairness with the Partially annotated Group labels (Fair-PG). We observe that the existing methods to achieve group fairness perform even worse than the vanilla training, which simply uses full data only with target labels, under Fair-PG. To address this problem, we propose a simple Confidence-based Group Label assignment (CGL) strategy that is readily applicable to any fairness-aware learning method. CGL utilizes an auxiliary group classifier to assign pseudo group labels, where random labels are assigned to low confident samples. We first theoretically show that our method design is better than the vanilla pseudo-labeling strategy in terms of fairness criteria. Then, we empirically show on several benchmark datasets that by combining CGL and the state-of-the-art fairness-aware in-processing methods, the target accuracies and the fairness metrics can be jointly improved compared to the baselines. Furthermore, we convincingly show that CGL enables to naturally augment the given group-labeled dataset with external target label-only datasets so that both accuracy and fairness can be improved. Code is available at https: //github.com/naver-ai/cgl_fairness.

Despite plenty of efforts focusing on improving the domain adaptation ability (DA) under unsupervised or few-shot semi-supervised settings, recently the solution of active learning started to attract more attention due to its suitability in transferring model in a more practical way with limited annotation resource on target data. Nevertheless, most active learning methods are not inherently designed to handle domain gap between data distribution, on the other hand, some active domain adaptation methods (ADA) usually requires complicated query functions, which is vulnerable to overfitting. In this work, we propose a concise but effective ADA method called Select-by-Distinctive-Margin (SDM), which consists of a maximum margin loss and a margin sampling algorithm for data selection. We provide theoretical analysis to show that SDM works like a Support Vector Machine, storing hard examples around decision boundaries and exploiting them to find informative and transferable data. In addition, we propose two variants of our method, one is designed to adaptively adjust the gradient from margin loss, the other boosts the selectivity of margin sampling by taking the gradient direction into account. We benchmark SDM with standard active learning setting, demonstrating our algorithm achieves competitive results with good data scalability.

Deep Implicit Function (DIF) has gained popularity as an efficient 3D shape representation. To capture geometry details, current methods usually learn DIF using local latent codes, which discretize the space into a regular 3D grid (or octree) and store local codes in grid points (or octree nodes). Given a query point, the local feature is computed by interpolating its neighboring local codes with their positions. However, the local codes are constrained at discrete and regular positions like grid points, which makes the code positions difficult to be optimized and limits their representation ability. To solve this problem, we propose to learn DIF with Dynamic Code Cloud, named DCC-DIF. Our method explicitly associates local codes with learnable position vectors, and the position vectors are continuous and can be dynamically optimized, which improves the representation ability. In addition, we propose a novel code position loss to optimize the code positions, which heuristically guides more local codes to be distributed around complex geometric details. In contrast to previous methods, our DCC-DIF represents 3D shapes more efficiently with a small amount of local codes, and improves the reconstruction quality. Experiments demonstrate that DCC-DIF achieves better performance over previous methods. Code and data are available at https://github.com/lity20/DCCDIF.

Canonical correlation analysis (CCA) matters in multiview representation learning. But, CCA and its most variants are essentially based on explicit or implicit covariance matrices. It means that they have no ability to model the nonlinear relationship among features due to intrinsic linearity of covariance. In this paper, we address the preceding problem and propose a novel canonical F-correlation framework by exploring and exploiting the nonlinear relationship between different features. The framework projects each feature rather than observation into a certain new space by an arbitrary nonlinear mapping, thus resulting in more flexibility in real applications. With this framework as a tool, we propose a correlative covariation projection (CCP) method by using an explicit nonlinear mapping. Moreover, we further propose a multiset version of CCP dubbed MCCP for learning compact representation of more than two views. The proposed MCCP is solved by an iterative method, and we prove the convergence of this iteration. A series of experimental results on six benchmark datasets demonstrate the effectiveness of our proposed CCP and MCCP methods.

Continual Learning (CL) methods aim to enable machine learning models to learn new tasks without catastrophic forgetting of those that have been previously mastered. Existing CL approaches often keep a buffer of previously-seen samples, perform knowledge distillation, or use regularization techniques towards this goal. Despite their performance, they still suffer from interference across tasks which leads to catastrophic forgetting. To ameliorate this problem, we propose to only activate and select sparse neurons for learning current and past tasks at any stage. More parameters space and model capacity can thus be reserved for the future tasks. This minimizes the interference between parameters for different tasks. To do so, we propose a Sparse neural Network for Continual Learning (SNCL), which employs variational Bayesian sparsity priors on the activations of the neurons in all layers. Full Experience Replay (FER) provides effective supervision in learning the sparse activations of the neurons in different layers. A loss-aware reservoir-sampling strategy is developed to maintain the memory buffer. The proposed method is agnostic as to the network structures and the task boundaries. Experiments on different datasets show that SNCL achieves state-of-the-art result for mitigating forgetting.

Multi-modality (i.e., multi-sensor) data is widely used in various vision tasks for more accurate or robust perception. However, the increased data modalities bring new challenges for data storage and transmission. The existing data compression approaches usually adopt individual codecs for each modality without considering the correlation between different modalities. This work proposes a multi-modality compression framework for infrared and visible image pairs by exploiting the cross-modality redundancy. Specifically, given the image in the reference modality (e.g., the infrared image), we use the channel-wise alignment module to produce the aligned features based on the affine transform. Then the aligned feature is used as the context information for compressing the image in the current modality (e.g., the visible image), and the corresponding affine coefficients are losslessly compressed at negligible cost. Furthermore, we introduce the Transformer-based spatial alignment module to exploit the correlation between the intermediate features in the decoding procedures for different modalities. Our framework is very flexible and easily extended for multi-modality video compression. Experimental results show our proposed framework outperforms the traditional and learning-based single modality compression methods on the FLIR and KAIST datasets.

Affordance grounding, a task to ground (i.e., localize) action possibility region in objects, which faces the challenge of establishing an explicit link with object parts due to the diversity of interactive affordance. Human has the ability that transform the various exocentric interactions to invariant egocentric affordance so as to counter the impact of interactive diversity. To empower an agent with such ability, this paper proposes a task of affordance grounding from exocentric view, i.e., given exocentric human-object interaction and egocentric object images, learning the affordance knowledge of the object and transferring it to the egocentric image using only the affordance label as supervision. To this end, we devise a cross-view knowledge transfer framework that extracts affordance-specific features from exocentric interactions and enhances the perception of affordance regions by preserving affordance correlation. Specifically, an Affordance Invariance Mining module is devised to extract specific clues by minimizing the intra-class differences originated from interaction habits in exocentric images. Besides, an Affordance Co-relation Preserving strategy is presented to perceive and localize affordance by aligning the co-relation matrix of predicted results between the two views. Particularly, an affordance grounding dataset named AGD20K is constructed by collecting and labeling over 20K images from 36 affordance categories. Experimental results demonstrate that our method outperforms the representative models in terms of objective metrics and visual quality. Code: github.com/lhc1224/Cross-View-AG.

Weakly-supervised semantic segmentation (WSSS) with image-level labels is an important and challenging task. Due to the high training efficiency, end-to-end solutions for WSSS have received increasing attention from the community. However, current methods are mainly based on convolutional neural networks and fail to explore the global information properly, thus usually resulting in incomplete object regions. In this paper, to address the aforementioned problem, we introduce Transformers, which naturally integrate global information, to generate more integral initial pseudo labels for end-to-end WSSS. Motivated by the inherent consistency between the self-attention in Transformers and the semantic affinity, we propose an Affinity from Attention (AFA) module to learn semantic affinity from the multi-head self-attention (MHSA) in Transformers. The learned affinity is then leveraged to refine the initial pseudo labels for segmentation. In addition, to efficiently derive reliable affinity labels for supervising AFA and ensure the local consistency of pseudo labels, we devise a Pixel-Adaptive Refinement module that incorporates low-level image appearance information to refine the pseudo labels. We perform extensive experiments and our method achieves 66.0% and 38.9% mIoU on the PASCAL VOC 2012 and MS COCO 2014 datasets, respectively, significantly outperforming recent end-to-end methods and several multi-stage competitors. Code is available at https://github.com/rulixiang/afa.

This paper proposes a real-world rolling shutter (RS) correction dataset, BS-RSC, and a corresponding model to correct the RS frames in a distorted video. Mobile devices in the consumer market with CMOS-based sensors for video capture often result in rolling shutter effects when relative movements occur during the video acquisition process, calling for RS effect removal techniques. However, current state-of-the-art RS correction methods often fail to remove RS effects in real scenarios since the motions are various and hard to model. To address this issue, we propose a real-world RS correction dataset BS-RSC. Real distorted videos with corresponding ground truth are recorded simultaneously via a well-designed beam-splitter-based acquisition system. BS-RSC contains various motions of both camera and objects in dynamic scenes. Further, an RS correction model with adaptive warping is proposed. Our model can warp the learned RS features into global shutter counterparts adaptively with predicted multiple displacement fields. These warped features are aggregated and then reconstructed into high-quality global shutter frames in a coarse-to-fine strategy. Experimental results demonstrate the effectiveness of the proposed method, and our dataset can improve the model's ability to remove the RS effects in the real world.

Representational learning forms the backbone of most deep learning applications, and the value of a learned representation is intimately tied to its information content regarding different factors of variation. Finding good representations depends on the nature of supervision and the learning algorithm. We propose a novel algorithm that utilizes a weak form of supervision where the data is partitioned into sets according to certain inactive (common) factors of variation which are invariant across elements of each set. Our key insight is that by seeking correspondence between elements of different sets, we learn strong representations that exclude the inactive factors of variation and isolate the active factors that vary within all sets. As a consequence of focusing on the active factors, our method can leverage a mix of set-supervised and wholly unsupervised data, which can even belong to a different domain. We tackle the challenging problem of synthetic-to-real object pose transfer, without pose annotations on anything, by isolating pose information which generalizes to the category level and across the synthetic/real domain gap. The method can also boost performance in supervised settings, by strengthening intermediate representations, as well as operate in practically attainable scenarios with set-supervised natural images, where quantity is limited and nuisance factors of variation are more plentiful.

A 3D scene consists of a set of objects, each with a shape and a layout giving their position in space. Understanding 3D scenes from 2D images is an important goal, with applications in robotics and graphics. While there have been recent advances in predicting 3D shape and layout from a single image, most approaches rely on 3D ground truth for training which is expensive to collect at scale. We overcome these limitations and propose a method that learns to predict 3D shape and layout for objects without any ground truth shape or layout information: instead we rely on multi-view images with 2D supervision which can more easily be collected at scale. Through extensive experiments on ShapeNet, Hypersim, and ScanNet we demonstrate that our approach scales to large datasets of realistic images, and compares favorably to methods relying on 3D ground truth. On Hypersim and ScanNet where reliable 3D ground truth is not available, our approach outperforms supervised approaches trained on smaller and less diverse datasets.

Vision Transformers (ViT) serve as powerful vision models. Unlike convolutional neural networks, which dominated vision research in previous years, vision transformers enjoy the ability to capture long-range dependencies in the data. Nonetheless, an integral part of any transformer architecture, the self-attention mechanism, suffers from high latency and inefficient memory utilization, making it less suitable for high-resolution input images. To alleviate these shortcomings, hierarchical vision models locally employ self-attention on non-interleaving windows. This relaxation reduces the complexity to be linear in the input size; however, it limits the cross-window interaction, hurting the model performance. In this paper, we propose a new shift-invariant local attention layer, called query and attend (QnA), that aggregates the input locally in an overlapping manner, much like convolutions. The key idea behind QnA is to introduce learned queries, which allow fast and efficient implementation. We verify the effectiveness of our layer by incorporating it into a hierarchical vision transformer model. We show improvements in speed and memory complexity while achieving comparable accuracy with state-of-the-art models. Finally, our layer scales especially well with window size, requiring up to x10 less memory while being up to x5 faster than existing methods. The code is publicly available at https://github.com/moabarar/qna.

Neural network quantization aims at reducing bit-widths of weights and activations for memory and computational efficiency. Since a linear quantizer (i.e., round(*) function) cannot well fit the bell-shaped distributions of weights and activations, many existing methods use pre-defined functions (e.g., exponential function) with learnable parameters to build the quantizer for joint optimization. However, these complicated quantizers introduce considerable computational overhead during inference since activation quantization should be conducted online. In this paper, we formulate the quantization process as a simple lookup operation and propose to learn lookup tables as quantizers. Specifically, we develop differentiable lookup tables and introduce several training strategies for optimization. Our lookup tables can be trained with the network in an end-to-end manner to fit the distributions in different layers and have very small additional computational cost. Comparison with previous methods show that quantized networks using our lookup tables achieve state-of-the-art performance on image classification, image super-resolution, and point cloud classification tasks.