CVPR 2022 Papers

Recent studies have shown the importance of modeling long-range interactions in the inpainting problem. To achieve this goal, existing approaches exploit either standalone attention techniques or transformers, but usually under a low resolution in consideration of computational cost. In this paper, we present a novel transformer-based model for large hole inpainting, which unifies the merits of transformers and convolutions to efficiently process high-resolution images. We carefully design each component of our framework to guarantee the high fidelity and diversity of recovered images. Specifically, we customize an inpainting-oriented transformer block, where the attention module aggregates non-local information only from partial valid tokens, indicated by a dynamic mask. Extensive experiments demonstrate the state-of-the-art performance of the new model on multiple benchmark datasets. Code is released at https://github.com/fenglinglwb/MAT.

The recent and increasing interest in video-language research has driven the development of large-scale datasets that enable data-intensive machine learning techniques. In comparison, limited effort has been made at assessing the fitness of these datasets for the video-language grounding task. Recent works have begun to discover significant limitations in these datasets, suggesting that state-of-the-art techniques commonly overfit to hidden dataset biases. In this work, we present MAD (Movie Audio Descriptions), a novel benchmark that departs from the paradigm of augmenting existing video datasets with text annotations and focuses on crawling and aligning available audio descriptions of mainstream movies. MAD contains over 384,000 natural language sentences grounded in over 1,200 hours of videos and exhibits a significant reduction in the currently diagnosed biases for video-language grounding datasets. MAD's collection strategy enables a novel and more challenging version of video-language grounding, where short temporal moments (typically seconds long) must be accurately grounded in diverse long-form videos that can last up to three hours. We have released MAD's data and baselines code at https://github.com/Soldelli/MAD.

Despite the potential of multi-modal pre-training to learn highly discriminative feature representations from complementary data modalities, current progress is being slowed by the lack of large-scale modality-diverse datasets. By leveraging the natural suitability of E-commerce, where different modalities capture complementary semantic information, we contribute a large-scale multi-modal pre-training dataset M5Product. The dataset comprises 5 modalities (image, text, table, video, and audio), covers over 6,000 categories and 5,000 attributes, and is 500 times larger than the largest publicly available dataset with a similar number of modalities. Furthermore, M5Product contains incomplete modality pairs and noise while also having a long-tailed distribution, resembling most real-world problems. We further propose Self-harmonized ContrAstive LEarning (SCALE), a novel pretraining framework that integrates the different modalities into a unified model through an adaptive feature fusion mechanism, where the importance of each modality is learned directly from the modality embeddings and impacts the inter-modality contrastive learning and masked tasks within a multi-modal transformer model. We evaluate the current multi-modal pre-training state-of-the-art approaches and benchmark their ability to learn from unlabeled data when faced with the large number of modalities in the M5Product dataset. We conduct extensive experiments on four downstream tasks and demonstrate the superiority of our SCALE model, providing insights into the importance of dataset scale and diversity.

In this study, we propose a three-dimensional Medical image classifier using Multi-plane and Multi-slice Transformer (M3T) network to classify Alzheimer's disease (AD) in 3D MRI images. The proposed network synergically combines 3D CNN, 2D CNN, and Transformer for accurate AD classification. The 3D CNN is used to perform natively 3D representation learning, while 2D CNN is used to utilize the pre-trained weights on large 2D databases and 2D representation learning. It is possible to efficiently extract the locality information for AD-related abnormalities in the local brain using CNN networks with inductive bias. The transformer network is also used to obtain attention relationships among multi-plane (axial, coronal, and sagittal) and multi-slice images after CNN. It is also possible to learn the abnormalities distributed over the wider region in the brain using the transformer without inductive bias. In this experiment, we used a training dataset from the Alzheimer's Disease Neuroimaging Initiative (ADNI) which contains a total of 4,786 3D T1-weighted MRI images. For the validation data, we used dataset from three different institutions: The Australian Imaging, Biomarker and Lifestyle Flagship Study of Ageing (AIBL), The Open Access Series of Imaging Studies (OASIS), and some set of ADNI data independent from the training dataset. Our proposed M3T is compared to conventional 3D classification networks based on an area under the curve (AUC) and classification accuracy for AD classification. This study represents that the proposed network M3T achieved the highest performance in multi-institutional validation database, and demonstrates the feasibility of the method to efficiently combine CNN and Transformer for 3D medical images.

Video editing tools are widely used nowadays for digital design. Although the demand for these tools is high, the prior knowledge required makes it difficult for novices to get started. Systems that could follow natural language instructions to perform automatic editing would significantly improve accessibility. This paper introduces the language-based video editing (LBVE) task, which allows the model to edit, guided by text instruction, a source video into a target video. LBVE contains two features: 1) the scenario of the source video is preserved instead of generating a completely different video; 2) the semantic is presented differently in the target video, and all changes are controlled by the given instruction. We propose a Multi-Modal Multi-Level Transformer (M3L-Transformer) to carry out LBVE. The M3L-Transformer dynamically learns the correspondence between video perception and language semantic at different levels, which benefits both the video understanding and video frame synthesis. We build three new datasets for evaluation, including two diagnostic and one from natural videos with human-labeled text. Extensive experimental results show that M3L-Transformer is effective for video editing and that LBVE can lead to a new field toward vision-and-language research.

Predicting future motions of road participants is an important task for driving autonomously in urban scenes. Existing models excel at predicting marginal trajectories for single agents, yet it remains an open question to jointly predict scene compliant trajectories over multiple agents. The challenge is due to exponentially increasing prediction space as a function of the number of agents. In this work, we exploit the underlying relations between interacting agents and decouple the joint prediction problem into marginal prediction problems. Our proposed approach M2I first classifies interacting agents as pairs of influencers and reactors, and then leverages a marginal prediction model and a conditional prediction model to predict trajectories for the influencers and reactors, respectively. The predictions from interacting agents are combined and selected according to their joint likelihoods. Experiments show that our simple but effective approach achieves state-of-the-art performance on the Waymo Open Motion Dataset interactive prediction benchmark.

Personalizing an avatar for co-speech gesture generation from spoken language requires learning the idiosyncrasies of a person's gesture style from a small amount of data. Previous methods in gesture generation require large amounts of data for each speaker, which is often infeasible. We propose an approach, named DiffGAN, that efficiently personalizes co-speech gesture generation models of a high-resource source speaker to target speaker with just 2 minutes of target training data. A unique characteristic of DiffGAN is its ability to account for the crossmodal grounding shift, while also addressing the distribution shift in the output domain. We substantiate the effectiveness of our approach a large scale publicly available dataset through quantitative, qualitative and user studies, which show that our proposed methodology significantly outperforms prior approaches for low-resource adaptation of gesture generation. Code and videos can be found at https://chahuja.com/diffgan

Human actions often induce changes of object states such as "cutting an apple", "cleaning shoes" or "pouring coffee". In this paper, we seek to temporally localize object states (e.g. "empty" and "full" cup) together with the corresponding state-modifying actions ("pouring coffee") in long uncurated videos with minimal supervision. The contributions of this work are threefold. First, we develop a self-supervised model for jointly learning state-modifying actions together with the corresponding object states from an uncurated set of videos from the Internet. The model is self-supervised by the causal ordering signal, i.e. initial object state -> manipulating action -> end state. Second, to cope with noisy uncurated training data, our model incorporates a noise adaptive weighting module supervised by a small number of annotated still images, that allows to efficiently filter out irrelevant videos during training. Third, we collect a new dataset with more than 2600 hours of video and 34 thousand changes of object states, and manually annotate a part of this data to validate our approach. Our results demonstrate substantial improvements over prior work in both action and object state-recognition in video.

Novel view synthesis from a single image has recently attracted a lot of attention, and it has been primarily advanced by 3D deep learning and rendering techniques. However, most work is still limited by synthesizing new views within relatively small camera motions. In this paper, we propose a novel approach to synthesize a consistent long-term video given a single scene image and a trajectory of large camera motions. Our approach utilizes an autoregressive Transformer to perform sequential modeling of multiple frames, which reasons the relations between multiple frames and the corresponding cameras to predict the next frame. To facilitate learning and ensure consistency among generated frames, we introduce a locality constraint based on the input cameras to guide self-attention among a large number of patches across space and time. Our method outperforms state-of-the-art view synthesis approaches by a large margin, especially when synthesizing long-term future in indoor 3D scenes. Project page at https://xrenaa.github.io/look-outside-room/.

Automatic font generation remains a challenging research issue due to the large amounts of characters with complicated structures. Typically, only a few samples can serve as the style/content reference (termed few-shot learning), which further increases the difficulty to preserve local style patterns or detailed glyph structures. We investigate the drawbacks of previous studies and find that a coarse-grained discriminator is insufficient for supervising a font generator. To this end, we propose a novel Component-Aware Module (CAM), which supervises the generator to decouple content and style at a more fine-grained level, i.e., the component level. Different from previous studies struggling to increase the complexity of generators, we aim to perform more effective supervision for a relatively simple generator to achieve its full potential, which is a brand new perspective for font generation. The whole framework achieves remarkable results by coupling component-level supervision with adversarial learning, hence we call it Component-Guided GAN, shortly CG-GAN. Extensive experiments show that our approach outperforms state-of-the-art one-shot font generation methods. Furthermore, it can be applied to handwritten word synthesis and scene text image editing, suggesting the generalization of our approach.

Temporal modeling is crucial for video super-resolution. Most of the video super-resolution methods adopt the optical flow or deformable convolution for explicitly motion compensation. However, such temporal modeling techniques increase the model complexity and might fail in case of occlusion or complex motion, resulting in serious distortion and artifacts. In this paper, we propose to explore the role of explicit temporal difference modeling in both LR and HR space. Instead of directly feeding consecutive frames into a VSR model, we propose to compute the temporal difference between frames and divide those pixels into two subsets according to the level of difference. They are separately processed with two branches of different receptive fields in order to better extract complementary information. To further enhance the super-resolution result, not only spatial residual features are extracted, but the difference between consecutive frames in high-frequency domain is also computed. It allows the model to exploit intermediate SR results in both future and past to refine the current SR output. The difference at different time steps could be cached such that information from further distance in time could be propagated to the current frame for refinement. Experiments on several video super-resolution benchmark datasets demonstrate the effectiveness of the proposed method and its favorable performance against state-of-the-art methods.

We address the problem of map sparsification for longterm visual localization. A commonly employed assumption in map sparsification is that the pre-build map and the later capture localization query are consistent. However, this assumption can be easily violated in the dynamic world. Additionally, the map size grows as new data accumulate through time, causing large data overhead in the long term. In this paper, we aim to overcome the environmental changes and reduce the map size at the same time by selecting points that are valuable to future localization. Inspired by the recent progress in Graph Neural Network (GNN), we propose the first work that models SfM maps as heterogeneous graphs and predicts 3D point importance scores with a GNN, which enables us to directly exploit the rich information in the SfM map graph. Two novel supervisions are proposed: 1) a data-fitting term for selecting valuable points to future localization based on training queries; 2) a K-Cover term for selecting sparse points with full-map coverage. In the experiments on a long-term dataset with environmental changes, our method selected map points on stable and widely visible structures and outperformed baselines in localization performance. This work novelly connects SfM maps with the abundant modern GNN techniques and opens a new research avenue forward.

Video frame interpolation (VFI) works generally predict intermediate frame(s) by first estimating the motion between inputs and then warping the inputs to the target time with the estimated motion. This approach, however, is not optimal when the temporal distance between the input sequence increases as existing motion estimation modules cannot effectively handle large motions. Hence, VFI works perform well for small frame gaps and perform poorly as the frame gap increases. In this work, we propose a novel framework to address this problem. We argue that when there is a large gap between inputs, instead of estimating imprecise motion that will eventually lead to inaccurate interpolation, we can safely propagate from one side of the input up to a reliable time frame using the other input as a reference. Then, the rest of the intermediate frames can be interpolated using standard approaches as the temporal gap is now narrowed. To this end, we propose a propagation network (PNet) by extending the classic feature-level forecasting with a novel motion-to-feature approach. To be thorough, we adopt a simple interpolation model along with PNet as our full model and design a simple procedure to train the full model in an end-to-end manner. Experimental results on several benchmark datasets confirm the effectiveness of our method for long-term VFI compared to state-of-the-art approaches.

Long-tailed data is still a big challenge for deep neural networks, even though they have achieved great success on balanced data. We observe that vanilla training on long-tailed data with cross-entropy loss makes the instance-rich head classes severely squeeze the spatial distribution of the tail classes, which leads to difficulty in classifying tail class samples. Furthermore, the original cross-entropy loss can only propagate gradient short-lively because the gradient in softmax form rapidly approaches zero as the logit difference increases. This phenomenon is called softmax saturation. It is unfavorable for training on balanced data, but can be utilized to adjust the validity of the samples in long-tailed data, thereby solving the distorted embedding space of long-tailed problems. To this end, this paper proposes the Gaussian clouded logit adjustment by Gaussian perturbation of different class logits with varied amplitude. We define the amplitude of perturbation as cloud size and set relatively large cloud sizes to tail classes. The large cloud size can reduce the softmax saturation and thereby making tail class samples more active as well as enlarging the embedding space. To alleviate the bias in a classifier, we therefore propose the class-based effective number sampling strategy with classifier re-training. Extensive experiments on benchmark datasets validate the superior performance of the proposed method. Source code is available at: https://github.com/Keke921/GCLLoss.

In the real open world, data tends to follow long-tailed class distributions, motivating the well-studied long-tailed recognition (LTR) problem. Naive training produces models that are biased toward common classes in terms of higher accuracy. The key to addressing LTR is to balance various aspects including data distribution, training losses, and gradients in learning. We explore an orthogonal direction, weight balancing , motivated by the empirical observation that the naively trained classifier has "artificially" larger weights in norm for common classes (because there exists abundant data to train them, unlike the rare classes). We investigate three techniques to balance weights, L2-normalization, weight decay, and MaxNorm. We first point out that L2-normalization "perfectly" balances per-class weights to be unit norm, but such a hard constraint might prevent classes from learning better classifiers. In contrast, weight decay penalizes larger weights more heavily and so learns small balanced weights; the MaxNorm constraint encourages growing small weights within a norm ball but caps all the weights by the radius. Our extensive study shows that both help learn balanced weights and greatly improve the LTR accuracy. Surprisingly, weight decay, although underexplored in LTR, significantly improves over prior work. Therefore, we adopt a two-stage training paradigm and propose a simple approach to LTR: (1) learning features using the cross-entropy loss by tuning weight decay, and (2) learning classifiers using class-balanced loss by tuning weight decay and MaxNorm. Our approach achieves the state-of-the-art accuracy on five standard benchmarks, serving as a future baseline for long-tailed recognition.

In this work, we introduce a novel strategy for long-tail recognition that addresses the tail classes' few-shot problem via training-free knowledge transfer. Our objective is to transfer knowledge acquired from information-rich common classes to semantically similar, and yet data-hungry, rare classes in order to obtain stronger tail class representations. We leverage the fact that class prototypes and learned cosine classifiers provide two different, complementary representations of class cluster centres in feature space, and use an attention mechanism to select and recompose learned classifiers features from common classes to obtain higher quality rare class representations. Our knowledge transfer process is training free, reducing overfitting risks, and can afford continual extension of classifiers to new classes. Experiments show that our approach can achieve significant performance boosts on rare classes while maintaining robust common class performance, outperforming directly comparable state-of-the-art models.

Video transformers have recently emerged as a competitive alternative to 3D CNNs for video understanding. However, due to their large number of parameters and reduced inductive biases, these models require supervised pretraining on large-scale image datasets to achieve top performance. In this paper, we empirically demonstrate that self-supervised pretraining of video transformers on video-only datasets can lead to action recognition results that are on par or better than those obtained with supervised pretraining on large-scale image datasets, even massive ones such as ImageNet-21K. Since transformer-based models are effective at capturing dependencies over extended temporal spans, we propose a simple learning procedure that forces the model to match a long-term view to a short-term view of the same video. Our approach, named Long-Short Temporal Contrastive Learning (LSTCL), enables video transformers to learn an effective clip-level representation by predicting temporal context captured from a longer temporal extent. To demonstrate the generality of our findings, we implement and validate our approach under three different self-supervised contrastive learning frameworks (MoCo v3, BYOL, SimSiam) using two distinct video-transformer architectures, including an improved variant of the Swin Transformer augmented with space-time attention. We conduct a thorough ablation study and show that LSTCL achieves competitive performance on multiple video benchmarks and represents a convincing alternative to supervised image-based pretraining.

Human pose estimation (HPE) usually requires large-scale training data to reach high performance. However, it is rather time-consuming to collect high-quality and fine-grained annotations for human body. To alleviate this issue, we revisit HPE and propose a location-free framework without supervision of keypoint locations. We reformulate the regression-based HPE from the perspective of classification. Inspired by the CAM-based weakly-supervised object localization, we observe that the coarse keypoint locations can be acquired through the part-aware CAMs but unsatisfactory due to the gap between the fine-grained HPE and the object-level localization. To this end, we propose a customized transformer framework to mine the fine-grained representation of human context, equipped with the structural relation to capture subtle differences among keypoints. Concretely, we design a Multi-scale Spatial-guided Context Encoder to fully capture the global human context while focusing on the part-aware regions and a Relation-encoded Pose Prototype Generation module to encode the structural relations. All these works together for strengthening the weak supervision from image-level category labels on locations. Our model achieves competitive performance on three datasets when only supervised at a category-level and importantly, it can achieve comparable results with fully-supervised methods with only 25% location labels on MS-COCO and MPII.

Deep learning methods can struggle to handle domain shifts not seen in training data, which can cause them to not generalize well to unseen domains. This has led to research attention on domain generalization (DG), which aims to the model's generalization ability to out-of-distribution. Adversarial domain generalization is a popular approach to DG, but conventional approaches (1) struggle to sufficiently align features so that local neighborhoods are mixed across domains; and (2) can suffer from feature space over collapse which can threaten generalization performance. To address these limitations, we propose localized adversarial domain generalization with space compactness maintenance (LADG) which constitutes two major contributions. First, we propose an adversarial localized classifier as the domain discriminator, along with a principled primary branch. This constructs a min-max game whereby the aim of the featurizer is to produce locally mixed domains. Second, we propose to use a coding-rate loss to alleviate feature space over collapse. We conduct comprehensive experiments on the Wilds DG benchmark to validate our approach, where LADG outperforms leading competitors on most datasets.