CVPR 2023 Papers

Large-scale vision-language models (VLMs) pre-trained on billion-level data have learned general visual representations and broad visual concepts. In principle, the well-learned knowledge structure of the VLMs should be inherited appropriately when being transferred to downstream tasks with limited data. However, most existing efficient transfer learning (ETL) approaches for VLMs either damage or are excessively biased towards the prior knowledge, e.g., prompt tuning (PT) discards the pre-trained text-based classifier and builds a new one while adapter-style tuning (AT) fully relies on the pre-trained features. To address this, we propose a new efficient tuning approach for VLMs named Task Residual Tuning (TaskRes), which performs directly on the text-based classifier and explicitly decouples the prior knowledge of the pre-trained models and new knowledge regarding a target task. Specifically, TaskRes keeps the original classifier weights from the VLMs frozen and obtains a new classifier for the target task by tuning a set of prior-independent parameters as a residual to the original one, which enables reliable prior knowledge preservation and flexible task-specific knowledge exploration. The proposed TaskRes is simple yet effective, which significantly outperforms previous ETL methods (e.g., PT and AT) on 11 benchmark datasets while requiring minimal effort for the implementation. Our code is available at https://github.com/geekyutao/TaskRes.

Vision-centric joint perception and prediction (PnP) has become an emerging trend in autonomous driving research. It predicts the future states of the traffic participants in the surrounding environment from raw RGB images. However, it is still a critical challenge to synchronize features obtained at multiple camera views and timestamps due to inevitable geometric distortions and further exploit those spatial-temporal features. To address this issue, we propose a temporal bird's-eye-view pyramid transformer (TBP-Former) for vision-centric PnP, which includes two novel designs. First, a pose-synchronized BEV encoder is proposed to map raw image inputs with any camera pose at any time to a shared and synchronized BEV space for better spatial-temporal synchronization. Second, a spatial-temporal pyramid transformer is introduced to comprehensively extract multi-scale BEV features and predict future BEV states with the support of spatial priors. Extensive experiments on nuScenes dataset show that our proposed framework overall outperforms all state-of-the-art vision-based prediction methods.

Vision Transformers (ViT) have shown competitive advantages in terms of performance compared to convolutional neural networks (CNNs), though they often come with high computational costs. To this end, previous methods explore different attention patterns by limiting a fixed number of spatially nearby tokens to accelerate the ViT's multi-head self-attention (MHSA) operations. However, such structured attention patterns limit the token-to-token connections to their spatial relevance, which disregards learned semantic connections from a full attention mask. In this work, we propose an approach to learn instance-dependent attention patterns, by devising a lightweight connectivity predictor module that estimates the connectivity score of each pair of tokens. Intuitively, two tokens have high connectivity scores if the features are considered relevant either spatially or semantically. As each token only attends to a small number of other tokens, the binarized connectivity masks are often very sparse by nature and therefore provide the opportunity to reduce network FLOPs via sparse computations. Equipped with the learned unstructured attention pattern, sparse attention ViT (Sparsifiner) produces a superior Pareto frontier between FLOPs and top-1 accuracy on ImageNet compared to token sparsity. Our method reduces 48% 69% FLOPs of MHSA while the accuracy drop is within 0.4%. We also show that combining attention and token sparsity reduces ViT FLOPs by over 60%.

High-resolution images enable neural networks to learn richer visual representations. However, this improved performance comes at the cost of growing computational complexity, hindering their usage in latency-sensitive applications. As not all pixels are equal, skipping computations for less-important regions offers a simple and effective measure to reduce the computation. This, however, is hard to be translated into actual speedup for CNNs since it breaks the regularity of the dense convolution workload. In this paper, we introduce SparseViT that revisits activation sparsity for recent window-based vision transformers (ViTs). As window attentions are naturally batched over blocks, actual speedup with window activation pruning becomes possible: i.e., 50% latency reduction with 60% sparsity. Different layers should be assigned with different pruning ratios due to their diverse sensitivities and computational costs. We introduce sparsity-aware adaptation and apply the evolutionary search to efficiently find the optimal layerwise sparsity configuration within the vast search space. SparseViT achieves speedups of 1.5x, 1.4x, and 1.3x compared to its dense counterpart in monocular 3D object detection, 2D instance segmentation, and 2D semantic segmentation, respectively, with negligible to no loss of accuracy.

Temporal Action Localization (TAL) methods typically operate on top of feature sequences from a frozen snippet encoder that is pretrained with the Trimmed Action Classification (TAC) tasks, resulting in a task discrepancy problem. While existing TAL methods mitigate this issue either by retraining the encoder with a pretext task or by end-to-end finetuning, they commonly require an overload of high memory and computation. In this work, we introduce Soft-Landing (SoLa) strategy, an efficient yet effective framework to bridge the transferability gap between the pretrained encoder and the downstream tasks by incorporating a light-weight neural network, i.e., a SoLa module, on top of the frozen encoder. We also propose an unsupervised training scheme for the SoLa module; it learns with inter-frame Similarity Matching that uses the frame interval as its supervisory signal, eliminating the need for temporal annotations. Experimental evaluation on various benchmarks for downstream TAL tasks shows that our method effectively alleviates the task discrepancy problem with remarkable computational efficiency.

Self-attention mechanism has been a key factor in the recent progress of Vision Transformer (ViT), which enables adaptive feature extraction from global contexts. However, existing self-attention methods either adopt sparse global attention or window attention to reduce the computation complexity, which may compromise the local feature learning or subject to some handcrafted designs. In contrast, local attention, which restricts the receptive field of each query to its own neighboring pixels, enjoys the benefits of both convolution and self-attention, namely local inductive bias and dynamic feature selection. Nevertheless, current local attention modules either use inefficient Im2Col function or rely on specific CUDA kernels that are hard to generalize to devices without CUDA support. In this paper, we propose a novel local attention module, Slide Attention, which leverages common convolution operations to achieve high efficiency, flexibility and generalizability. Specifically, we first re-interpret the column-based Im2Col function from a new row-based perspective and use Depthwise Convolution as an efficient substitution. On this basis, we propose a deformed shifting module based on the re-parameterization technique, which further relaxes the fixed key/value positions to deformed features in the local region. In this way, our module realizes the local attention paradigm in both efficient and flexible manner. Extensive experiments show that our slide attention module is applicable to a variety of advanced Vision Transformer models and compatible with various hardware devices, and achieves consistently improved performances on comprehensive benchmarks.

Cross-modal alignment is essential for vision-language pre-training (VLP) models to learn the correct corresponding information across different modalities. For this purpose, inspired by the success of masked language modeling (MLM) tasks in the NLP pre-training area, numerous masked modeling tasks have been proposed for VLP to further promote cross-modal interactions. The core idea of previous masked modeling tasks is to focus on reconstructing the masked tokens based on visible context for learning local-to-local alignment. However, most of them pay little attention to the global semantic features generated for the masked data, resulting in a limited cross-modal alignment ability of global representations. Therefore, in this paper, we propose a novel Semantic Completion Learning (SCL) task, complementary to existing masked modeling tasks, to facilitate global-to-local alignment. Specifically, the SCL task complements the missing semantics of masked data by capturing the corresponding information from the other modality, promoting learning more representative global features which have a great impact on the performance of downstream tasks. Moreover, we present a flexible vision encoder, which enables our model to perform image-text and video-text multimodal tasks simultaneously. Experimental results show that our proposed method obtains state-of-the-art performance on various vision-language benchmarks, such as visual question answering, image-text retrieval, and video-text retrieval.

Despite the success of deep learning in video understanding tasks, processing every frame in a video is computationally expensive and often unnecessary in real-time applications. Frame selection aims to extract the most informative and representative frames to help a model better understand video content. Existing frame selection methods either individually sample frames based on per-frame importance prediction, without considering interaction among frames, or adopt reinforcement learning agents to find representative frames in succession, which are costly to train and may lead to potential stability issues. To overcome the limitations of existing methods, we propose a Search-Map-Search learning paradigm which combines the advantages of heuristic search and supervised learning to select the best combination of frames from a video as one entity. By combining search with learning, the proposed method can better capture frame interactions while incurring a low inference overhead. Specifically, we first propose a hierarchical search method conducted on each training video to search for the optimal combination of frames with the lowest error on the downstream task. A feature mapping function is then learned to map the frames of a video to the representation of its target optimal frame combination. During inference, another search is performed on an unseen video to select a combination of frames whose feature representation is close to the projected feature representation. Extensive experiments based on several action recognition benchmarks demonstrate that our frame selection method effectively improves performance of action recognition models, and significantly outperforms a number of competitive baselines.

Semi-supervised action recognition is a challenging but critical task due to the high cost of video annotations. Existing approaches mainly use convolutional neural networks, yet current revolutionary vision transformer models have been less explored. In this paper, we investigate the use of transformer models under the SSL setting for action recognition. To this end, we introduce SVFormer, which adopts a steady pseudo-labeling framework (ie, EMA-Teacher) to cope with unlabeled video samples. While a wide range of data augmentations have been shown effective for semi-supervised image classification, they generally produce limited results for video recognition. We therefore introduce a novel augmentation strategy, Tube TokenMix, tailored for video data where video clips are mixed via a mask with consistent masked tokens over the temporal axis. In addition, we propose a temporal warping augmentation to cover the complex temporal variation in videos, which stretches selected frames to various temporal durations in the clip. Extensive experiments on three datasets Kinetics-400, UCF-101, and HMDB-51 verify the advantage of SVFormer. In particular, SVFormer outperforms the state-of-the-art by 31.5% with fewer training epochs under the 1% labeling rate of Kinetics-400. Our method can hopefully serve as a strong benchmark and encourage future search on semi-supervised action recognition with Transformer networks.

Traditional video action detectors typically adopt the two-stage pipeline, where a person detector is first employed to yield actor boxes and then 3D RoIAlign is used to extract actor-specific features for classification. This detection paradigm requires multi-stage training and inference and cannot capture context information outside the bounding box. Recently, a few query-based action detectors are proposed to predict action instances in an end-to-end manner. However, they still lack adaptability in feature sampling or decoding, thus suffering from the issue of inferior performance or slower convergence. In this paper, we propose a new one-stage sparse action detector, termed STMixer. STMixer is based on two core designs. First, we present a query-based adaptive feature sampling module, which endows our STMixer with the flexibility of mining a set of discriminative features from the entire spatiotemporal domain. Second, we devise a dual-branch feature mixing module, which allows our STMixer to dynamically attend to and mix video features along the spatial and the temporal dimension respectively for better feature decoding. Coupling these two designs with a video backbone yields an efficient and accurate action detector. Without bells and whistles, STMixer obtains the state-of-the-art results on the datasets of AVA, UCF101-24, and JHMDB.

We study the problem of human action recognition using motion capture (MoCap) sequences. Unlike existing techniques that take multiple manual steps to derive standardized skeleton representations as model input, we propose a novel Spatial-Temporal Mesh Transformer (STMT) to directly model the mesh sequences. The model uses a hierarchical transformer with intra-frame off-set attention and inter-frame self-attention. The attention mechanism allows the model to freely attend between any two vertex patches to learn non-local relationships in the spatial-temporal domain. Masked vertex modeling and future frame prediction are used as two self-supervised tasks to fully activate the bi-directional and auto-regressive attention in our hierarchical transformer. The proposed method achieves state-of-the-art performance compared to skeleton-based and point-cloud-based models on common MoCap benchmarks. Code is available at https://github.com/zgzxy001/STMT.

Timestamp supervised temporal action segmentation (TSTAS) is more cost-effective than fully supervised counterparts. However, previous approaches suffer from severe label bias due to over-reliance on sparse timestamp annotations, resulting in unsatisfactory performance. In this paper, we propose the Debiasing-TSTAS (D-TSTAS) framework by exploiting unannotated frames to alleviate this bias from two phases: 1) Initialization. To reduce the dependencies on annotated frames, we propose masked timestamp predictions (MTP) to ensure that initialized model captures more contextual information. 2) Refinement. To overcome the limitation of the expressiveness from sparsely annotated timestamps, we propose a center-oriented timestamp expansion (CTE) approach to progressively expand pseudo-timestamp groups which contain semantic-rich motion representation of action segments. Then, these pseudo-timestamp groups and the model output are used to iteratively generate pseudo-labels for refining the model in a fully supervised setup. We further introduce segmental confidence loss to enable the model to have high confidence predictions within the pseudo-timestamp groups and more accurate action boundaries. Our D-TSTAS outperforms the state-of-the-art TSTAS method as well as achieves competitive results compared with fully supervised approaches on three benchmark datasets.

The large-scale deployment of autonomous vehicles is yet to come, and one of the major remaining challenges lies in urban dense traffic scenarios. In such cases, it remains challenging to predict the future evolution of the scene and future behaviors of objects, and to deal with rare adverse events such as the sudden appearance of occluded objects. In this paper, we present ReasonNet, a novel end-to-end driving framework that extensively exploits both temporal and global information of the driving scene. By reasoning on the temporal behavior of objects, our method can effectively process the interactions and relationships among features in different frames. Reasoning about the global information of the scene can also improve overall perception performance and benefit the detection of adverse events, especially the anticipation of potential danger from occluded objects. For comprehensive evaluation on occlusion events, we also release publicly a driving simulation benchmark DriveOcclusionSim consisting of diverse occlusion events. We conduct extensive experiments on multiple CARLA benchmarks, where our model outperforms all prior methods, ranking first on the sensor track of the public CARLA Leaderboard.

Temporal action localization (TAL) requires long-form reasoning to predict actions of various durations and complex content. Given limited GPU memory, training TAL end to end (i.e., from videos to predictions) on long videos is a significant challenge. Most methods can only train on pre-extracted features without optimizing them for the localization problem, consequently limiting localization performance. In this work, to extend the potential in TAL networks, we propose a novel end-to-end method Re2TAL, which rewires pretrained video backbones for reversible TAL. Re2TAL builds a backbone with reversible modules, where the input can be recovered from the output such that the bulky intermediate activations can be cleared from memory during training. Instead of designing one single type of reversible module, we propose a network rewiring mechanism, to transform any module with a residual connection to a reversible module without changing any parameters. This provides two benefits: (1) a large variety of reversible networks are easily obtained from existing and even future model designs, and (2) the reversible models require much less training effort as they reuse the pre-trained parameters of their original non-reversible versions. Re2TAL, only using the RGB modality, reaches 37.01% average mAP on ActivityNet-v1.3, a new state-of-the-art record, and mAP 64.9% at tIoU=0.5 on THUMOS-14, outperforming all other RGB-only methods. Code is available at https://github.com/coolbay/Re2TAL.

This paper studies how to keep a vision backbone effective while removing token mixers in its basic building blocks. Token mixers, as self-attention for vision transformers (ViTs), are intended to perform information communication between different spatial tokens but suffer from considerable computational cost and latency. However, directly removing them will lead to an incomplete model structure prior, and thus brings a significant accuracy drop. To this end, we first develop an RepIdentityFormer base on the re-parameterizing idea, to study the token mixer free model architecture. And we then explore the improved learning paradigm to break the limitation of simple token mixer free backbone, and summarize the empirical practice into 5 guidelines. Equipped with the proposed optimization strategy, we are able to build an extremely simple vision backbone with encouraging performance, while enjoying the high efficiency during inference. Extensive experiments and ablative analysis also demonstrate that the inductive bias of network architecture, can be incorporated into simple network structure with appropriate optimization strategy. We hope this work can serve as a starting point for the exploration of optimization-driven efficient network design.

Most neural networks for computer vision are designed to infer using RGB images. However, these RGB images are commonly encoded in JPEG before saving to disk; decoding them imposes an unavoidable overhead for RGB networks. Instead, our work focuses on training Vision Transformers (ViT) directly from the encoded features of JPEG. This way, we can avoid most of the decoding overhead, accelerating data load. Existing works have studied this aspect but they focus on CNNs. Due to how these encoded features are structured, CNNs require heavy modification to their architecture to accept such data. Here, we show that this is not the case for ViTs. In addition, we tackle data augmentation directly on these encoded features, which to our knowledge, has not been explored in-depth for training in this setting. With these two improvements -- ViT and data augmentation -- we show that our ViT-Ti model achieves up to 39.2% faster training and 17.9% faster inference with no accuracy loss compared to the RGB counterpart.

Finetuning a large vision language model (VLM) on a target dataset after large scale pretraining is a dominant paradigm in visual question answering (VQA). Datasets for specialized tasks such as knowledge-based VQA or VQA in non natural-image domains are orders of magnitude smaller than those for general-purpose VQA. While collecting additional labels for specialized tasks or domains can be challenging, unlabeled images are often available. We introduce SelTDA (Self-Taught Data Augmentation), a strategy for finetuning large VLMs on small-scale VQA datasets. SelTDA uses the VLM and target dataset to build a teacher model that can generate question-answer pseudolabels directly conditioned on an image alone, allowing us to pseudolabel unlabeled images. SelTDA then finetunes the initial VLM on the original dataset augmented with freshly pseudolabeled images. We describe a series of experiments showing that our self-taught data augmentation increases robustness to adversarially searched questions, counterfactual examples, and rephrasings, it improves domain generalization, and results in greater retention of numerical reasoning skills. The proposed strategy requires no additional annotations or architectural modifications, and is compatible with any modern encoder-decoder multimodal transformer. Code available at https://github.com/codezakh/SelTDA

Weakly-supervised temporal action localization aims to localize and recognize actions in untrimmed videos with only video-level category labels during training. Without instance-level annotations, most existing methods follow the Segment-based Multiple Instance Learning (S-MIL) framework, where the predictions of segments are supervised by the labels of videos. However, the objective for acquiring segment-level scores during training is not consistent with the target for acquiring proposal-level scores during testing, leading to suboptimal results. To deal with this problem, we propose a novel Proposal-based Multiple Instance Learning (P-MIL) framework that directly classifies the candidate proposals in both the training and testing stages, which includes three key designs: 1) a surrounding contrastive feature extraction module to suppress the discriminative short proposals by considering the surrounding contrastive information, 2) a proposal completeness evaluation module to inhibit the low-quality proposals with the guidance of the completeness pseudo labels, and 3) an instance-level rank consistency loss to achieve robust detection by leveraging the complementarity of RGB and FLOW modalities. Extensive experimental results on two challenging benchmarks including THUMOS14 and ActivityNet demonstrate the superior performance of our method. Our code is available at github.com/OpenSpaceAI/CVPR2023_P-MIL.

This study investigates unsupervised anomaly action recognition, which identifies video-level abnormal-human-behavior events in an unsupervised manner without abnormal samples, and simultaneously addresses three limitations in the conventional skeleton-based approaches: target domain-dependent DNN training, robustness against skeleton errors, and a lack of normal samples. We present a unified, user prompt-guided zero-shot learning framework using a target domain-independent skeleton feature extractor, which is pretrained on a large-scale action recognition dataset. Particularly, during the training phase using normal samples, the method models the distribution of skeleton features of the normal actions while freezing the weights of the DNNs and estimates the anomaly score using this distribution in the inference phase. Additionally, to increase robustness against skeleton errors, we introduce a DNN architecture inspired by a point cloud deep learning paradigm, which sparsely propagates the features between joints. Furthermore, to prevent the unobserved normal actions from being misidentified as abnormal actions, we incorporate a similarity score between the user prompt embeddings and skeleton features aligned in the common space into the anomaly score, which indirectly supplements normal actions. On two publicly available datasets, we conduct experiments to test the effectiveness of the proposed method with respect to abovementioned limitations.