CVPR 2023 论文列表

Vision-Language Pre-Training (VLP) has shown promising capabilities to align image and text pairs, facilitating a broad variety of cross-modal learning tasks. However, we observe that VLP models often lack the visual grounding/localization capability which is critical for many downstream tasks such as visual reasoning. In this work, we propose a novel Position-guided Text Prompt (PTP) paradigm to enhance the visual grounding ability of cross-modal models trained with VLP. Specifically, in the VLP phase, PTP divides the image into NxN blocks, and identifies the objects in each block through the widely used object detector in VLP. It then reformulates the visual grounding task into a fill-in-the-blank problem given a PTP by encouraging the model to predict the objects in the given blocks or regress the blocks of a given object, e.g. filling "P" or "O" in a PTP "The block P has a O". This mechanism improves the visual grounding capability of VLP models and thus helps them better handle various downstream tasks. By introducing PTP into several state-of-the-art VLP frameworks, we observe consistently significant improvements across representative cross-modal learning model architectures and several benchmarks, e.g. zero-shot Flickr30K Retrieval (+4.8 in average recall@1) for ViLT baseline, and COCO Captioning (+5.3 in CIDEr) for SOTA BLIP baseline. Moreover, PTP achieves comparable results with object-detector based methods, and much faster inference speed since PTP discards its object detector for inference while the later cannot. Our code and pre-trained weight will be released.

Weakly-supervised Temporal Action Localization (WTAL) attempts to localize the actions in untrimmed videos using only video-level supervision. Most recent works approach WTAL from a localization-by-classification perspective where these methods try to classify each video frame followed by a manually-designed post-processing pipeline to aggregate these per-frame action predictions into action snippets. Due to this perspective, the model lacks any explicit understanding of action boundaries and tends to focus only on the most discriminative parts of the video resulting in incomplete action localization. To address this, we present PivoTAL, Prior-driven Supervision for Weakly-supervised Temporal Action Localization, to approach WTAL from a localization-by-localization perspective by learning to localize the action snippets directly. To this end, PivoTAL leverages the underlying spatio-temporal regularities in videos in the form of action-specific scene prior, action snippet generation prior, and learnable Gaussian prior to supervise the localization-based training. PivoTAL shows significant improvement (of at least 3% avg mAP) over all existing methods on the benchmark datasets, THUMOS-14 and ActivitNet-v1.3.

Vision Transformers (ViTs) are built on the assumption of treating image patches as "visual tokens" and learn patch-to-patch attention. The patch embedding based tokenizer has a semantic gap with respect to its counterpart, the textual tokenizer. The patch-to-patch attention suffers from the quadratic complexity issue, and also makes it non-trivial to explain learned ViTs. To address these issues in ViT, this paper proposes to learn Patch-to-Cluster attention (PaCa) in ViT. Queries in our PaCa-ViT starts with patches, while keys and values are directly based on clustering (with a predefined small number of clusters). The clusters are learned end-to-end, leading to better tokenizers and inducing joint clustering-for-attention and attention-for-clustering for better and interpretable models. The quadratic complexity is relaxed to linear complexity. The proposed PaCa module is used in designing efficient and interpretable ViT backbones and semantic segmentation head networks. In experiments, the proposed methods are tested on ImageNet-1k image classification, MS-COCO object detection and instance segmentation and MIT-ADE20k semantic segmentation. Compared with the prior art, it obtains better performance in all the three benchmarks than the SWin and the PVTs by significant margins in ImageNet-1k and MIT-ADE20k. It is also significantly more efficient than PVT models in MS-COCO and MIT-ADE20k due to the linear complexity. The learned clusters are semantically meaningful. Code and model checkpoints are available at https://github.com/iVMCL/PaCaViT.

Existing methods for open set action recognition focus on novelty detection that assumes video clips show a single action, which is unrealistic in the real world. We propose a new method for open set action recognition and novelty detection via MUlti-Label Evidential learning (MULE), that goes beyond previous novel action detection methods by addressing the more general problems of single or multiple actors in the same scene, with simultaneous action(s) by any actor. Our Beta Evidential Neural Network estimates multi-action uncertainty with Beta densities based on actor-context-object relation representations. An evidence debiasing constraint is added to the objective func- tion for optimization to reduce the static bias of video representations, which can incorrectly correlate predictions and static cues. We develop a primal-dual average scheme update-based learning algorithm to optimize the proposed problem and provide corresponding theoretical analysis. Besides, uncertainty and belief-based novelty estimation mechanisms are formulated to detect novel actions. Extensive experiments on two real-world video datasets show that our proposed approach achieves promising performance in single/multi-actor, single/multi-action settings. Our code and models are released at https://github.com/charliezhaoyinpeng/mule.

Learning-based methods to solve dense 3D vision problems typically train on 3D sensor data. The respectively used principle of measuring distances provides advantages and drawbacks. These are typically not compared nor discussed in the literature due to a lack of multi-modal datasets. Texture-less regions are problematic for structure from motion and stereo, reflective material poses issues for active sensing, and distances for translucent objects are intricate to measure with existing hardware. Training on inaccurate or corrupt data induces model bias and hampers generalisation capabilities. These effects remain unnoticed if the sensor measurement is considered as ground truth during the evaluation. This paper investigates the effect of sensor errors for the dense 3D vision tasks of depth estimation and reconstruction. We rigorously show the significant impact of sensor characteristics on the learned predictions and notice generalisation issues arising from various technologies in everyday household environments. For evaluation, we introduce a carefully designed dataset comprising measurements from commodity sensors, namely D-ToF, I-ToF, passive/active stereo, and monocular RGB+P. Our study quantifies the considerable sensor noise impact and paves the way to improved dense vision estimates and targeted data fusion.

In this work we study the benefits of using tracking and 3D poses for action recognition. To achieve this, we take the Lagrangian view on analysing actions over a trajectory of human motion rather than at a fixed point in space. Taking this stand allows us to use the tracklets of people to predict their actions. In this spirit, first we show the benefits of using 3D pose to infer actions, and study person-person interactions. Subsequently, we propose a Lagrangian Action Recognition model by fusing 3D pose and contextualized appearance over tracklets. To this end, our method achieves state-of-the-art performance on the AVA v2.2 dataset on both pose only settings and on standard benchmark settings. When reasoning about the action using only pose cues, our pose model achieves +10.0 mAP gain over the corresponding state-of-the-art while our fused model has a gain of +2.8 mAP over the best state-of-the-art model. Code and results are available at: https://brjathu.github.io/LART

The complicated architecture and high training cost of vision transformers urge the exploration of post-training quantization. However, the heavy-tailed distribution of vision transformer activations hinders the effectiveness of previous post-training quantization methods, even with advanced quantizer designs. Instead of tuning the quantizer to better fit the complicated activation distribution, this paper proposes NoisyQuant, a quantizer-agnostic enhancement for the post-training activation quantization performance of vision transformers. We make a surprising theoretical discovery that for a given quantizer, adding a fixed Uniform noisy bias to the values being quantized can significantly reduce the quantization error under provable conditions. Building on the theoretical insight, NoisyQuant achieves the first success on actively altering the heavy-tailed activation distribution with additive noisy bias to fit a given quantizer. Extensive experiments show NoisyQuant largely improves the post-training quantization performance of vision transformer with minimal computation overhead. For instance, on linear uniform 6-bit activation quantization, NoisyQuant improves SOTA top-1 accuracy on ImageNet by up to 1.7%, 1.1% and 0.5% for ViT, DeiT, and Swin Transformer respectively, achieving on-par or even higher performance than previous nonlinear, mixed-precision quantization.

Training a Neural Radiance Field (NeRF) without pre-computed camera poses is challenging. Recent advances in this direction demonstrate the possibility of jointly optimising a NeRF and camera poses in forward-facing scenes. However, these methods still face difficulties during dramatic camera movement. We tackle this challenging problem by incorporating undistorted monocular depth priors. These priors are generated by correcting scale and shift parameters during training, with which we are then able to constrain the relative poses between consecutive frames. This constraint is achieved using our proposed novel loss functions. Experiments on real-world indoor and outdoor scenes show that our method can handle challenging camera trajectories and outperforms existing methods in terms of novel view rendering quality and pose estimation accuracy. Our project page is https://nope-nerf.active.vision.

High-definition (HD) semantic maps are a crucial component for autonomous driving on urban streets. Traditional offline HD maps are created through labor-intensive manual annotation processes, which are costly and do not accommodate timely updates. Recently, researchers have proposed to infer local maps based on online sensor observations. However, the range of online map inference is constrained by sensor perception range and is easily affected by occlusions. In this work, we propose Neural Map Prior (NMP), a neural representation of global maps that enables automatic global map updates and enhances local map inference performance. To incorporate the strong map prior into local map inference, we leverage cross-attention to dynamically capture the correlations between current features and prior features. For updating the global neural map prior, we use a learning-based fusion module to guide the network in fusing features from previous traversals. This design allows the network to capture a global neural map prior while making sequential online map predictions. Experimental results on the nuScenes dataset demonstrate that our framework is compatible with most map segmentation/detection methods, improving map prediction performance in challenging weather conditions and over an extended horizon. To the best of our knowledge, this represents the first learning-based system for constructing a global map prior.

Skeleton-based human action recognition is becoming increasingly important in a variety of fields. Most existing works train a CNN or GCN based backbone to extract spatial-temporal features, and use temporal average/max pooling to aggregate the information. However, these pooling methods fail to capture high-order dynamics information. To address the problem, we propose a plug-and-play module called Koopman pooling, which is a parameterized high-order pooling technique based on Koopman theory. The Koopman operator linearizes a non-linear dynamics system, thus providing a way to represent the complex system through the dynamics matrix, which can be used for classification. We also propose an eigenvalue normalization method to encourage the learned dynamics to be non-decaying and stable. Besides, we also show that our Koopman pooling framework can be easily extended to one-shot action recognition when combined with Dynamic Mode Decomposition. The proposed method is evaluated on three benchmark datasets, namely NTU RGB+D 60, 120 and NW-UCLA. Our experiments clearly demonstrate that Koopman pooling significantly improves the performance under both full-dataset and one-shot settings.

The task of reconstructing 3D human motion has wide-ranging applications. The gold standard Motion capture (MoCap) systems are accurate but inaccessible to the general public due to their cost, hardware, and space constraints. In contrast, monocular human mesh recovery (HMR) methods are much more accessible than MoCap as they take single-view videos as inputs. Replacing the multi-view MoCap systems with a monocular HMR method would break the current barriers to collecting accurate 3D motion thus making exciting applications like motion analysis and motion-driven animation accessible to the general public. However, the performance of existing HMR methods degrades when the video contains challenging and dynamic motion that is not in existing MoCap datasets used for training. This reduces its appeal as dynamic motion is frequently the target in 3D motion recovery in the aforementioned applications. Our study aims to bridge the gap between monocular HMR and multi-view MoCap systems by leveraging information shared across multiple video instances of the same action. We introduce the Neural Motion (NeMo) field. It is optimized to represent the underlying 3D motions across a set of videos of the same action. Empirically, we show that NeMo can recover 3D motion in sports using videos from the Penn Action dataset, where NeMo outperforms existing HMR methods in terms of 2D keypoint detection. To further validate NeMo using 3D metrics, we collected a small MoCap dataset mimicking actions in Penn Action, and show that NeMo achieves better 3D reconstruction compared to various baselines.

The main challenge in vision-and-language navigation (VLN) is how to understand natural-language instructions in an unseen environment. The main limitation of conventional VLN algorithms is that if an action is mistaken, the agent fails to follow the instructions or explores unnecessary regions, leading the agent to an irrecoverable path. To tackle this problem, we propose Meta-Explore, a hierarchical navigation method deploying an exploitation policy to correct misled recent actions. We show that an exploitation policy, which moves the agent toward a well-chosen local goal among unvisited but observable states, outperforms a method which moves the agent to a previously visited state. We also highlight the demand for imagining regretful explorations with semantically meaningful clues. The key to our approach is understanding the object placements around the agent in spectral-domain. Specifically, we present a novel visual representation, called scene object spectrum (SOS), which performs category-wise 2D Fourier transform of detected objects. Combining exploitation policy and SOS features, the agent can correct its path by choosing a promising local goal. We evaluate our method in three VLN benchmarks: R2R, SOON, and REVERIE. Meta-Explore outperforms other baselines and shows significant generalization performance. In addition, local goal search using the proposed spectral-domain SOS features significantly improves the success rate by 17.1% and SPL by 20.6% for the SOON benchmark.

Position Embeddings (PEs), an arguably indispensable component in Vision Transformers (ViTs), have been shown to improve the performance of ViTs on many vision tasks. However, PEs have a potentially high risk of privacy leakage since the spatial information of the input patches is exposed. This caveat naturally raises a series of interesting questions about the impact of PEs on accuracy, privacy, prediction consistency, etc. To tackle these issues, we propose a Masked Jigsaw Puzzle (MJP) position embedding method. In particular, MJP first shuffles the selected patches via our block-wise random jigsaw puzzle shuffle algorithm, and their corresponding PEs are occluded. Meanwhile, for the non-occluded patches, the PEs remain the original ones but their spatial relation is strengthened via our dense absolute localization regressor. The experimental results reveal that 1) PEs explicitly encode the 2D spatial relationship and lead to severe privacy leakage problems under gradient inversion attack; 2) Training ViTs with the naively shuffled patches can alleviate the problem, but it harms the accuracy; 3) Under a certain shuffle ratio, the proposed MJP not only boosts the performance and robustness on large-scale datasets (i.e., ImageNet-1K and ImageNet-C, -A/O) but also improves the privacy preservation ability under typical gradient attacks by a large margin. The source code and trained models are available at https://github.com/yhlleo/MJP.

In this paper, we study a novel problem in egocentric action recognition, which we term as "Multimodal Generalization" (MMG). MMG aims to study how systems can generalize when data from certain modalities is limited or even completely missing. We thoroughly investigate MMG in the context of standard supervised action recognition and the more challenging few-shot setting for learning new action categories. MMG consists of two novel scenarios, designed to support security, and efficiency considerations in real-world applications: (1) missing modality generalization where some modalities that were present during the train time are missing during the inference time, and (2) cross-modal zero-shot generalization, where the modalities present during the inference time and the training time are disjoint. To enable this investigation, we construct a new dataset MMG-Ego4D containing data points with video, audio, and inertial motion sensor (IMU) modalities. Our dataset is derived from Ego4D dataset, but processed and thoroughly re-annotated by human experts to facilitate research in the MMG problem. We evaluate a diverse array of models on MMG-Ego4D and propose new methods with improved generalization ability. In particular, we introduce a new fusion module with modality dropout training, contrastive-based alignment training, and a novel cross-modal prototypical loss for better few-shot performance. We hope this study will serve as a benchmark and guide future research in multimodal generalization problems. The benchmark and code are available at https://github.com/facebookresearch/MMG_Ego4D

Multimodal semantic understanding often has to deal with uncertainty, which means the obtained messages tend to refer to multiple targets. Such uncertainty is problematic for our interpretation, including inter- and intra-modal uncertainty. Little effort has studied the modeling of this uncertainty, particularly in pre-training on unlabeled datasets and fine-tuning in task-specific downstream datasets. In this paper, we project the representations of all modalities as probabilistic distributions via a Probability Distribution Encoder (PDE) by utilizing sequence-level interactions. Compared to the exiting deterministic methods, such uncertainty modeling can convey richer multimodal semantic information and more complex relationships. Furthermore, we integrate uncertainty modeling with popular pre-training frameworks and propose suitable pre-training tasks: Distribution-based Vision-Language Contrastive learning (D-VLC), Distribution-based Masked Language Modeling (D-MLM), and Distribution-based Image-Text Matching (D-ITM). The fine-tuned models are applied to challenging downstream tasks, including image-text retrieval, visual question answering, visual reasoning, and visual entailment, and achieve state-of-the-art results.

Most existing vision-language pre-training (VLP) approaches adopt cross-modal masked language modeling (CMLM) to learn vision-language associations. However, we find that CMLM is insufficient for this purpose according to our observations: (1) Modality bias: a considerable amount of masked tokens in CMLM can be recovered with only the language information, ignoring the visual inputs. (2) Under-utilization of the unmasked tokens: CMLM primarily focuses on the masked token but it cannot simultaneously leverage other tokens to learn vision-language associations. To handle those limitations, we propose EPIC (lEveraging Per Image-Token Consistency for vision-language pre-training). In EPIC, for each image-sentence pair, we mask tokens that are salient to the image (i.e., Saliency-based Masking Strategy) and replace them with alternatives sampled from a language model (i.e., Inconsistent Token Generation Procedure), and then the model is required to determine for each token in the sentence whether it is consistent with the image (i.e., Image-Token Consistency Task). The proposed EPIC method is easily combined with pre-training methods. Extensive experiments show that the combination of the EPIC method and state-of-the-art pre-training approaches, including ViLT, ALBEF, METER, and X-VLM, leads to significant improvements on downstream tasks. Our coude is released at https://github.com/gyhdog99/epic

Large scale vision and language models can achieve impressive zero-shot recognition performance by mapping class specific text queries to image content. Two distinct challenges that remain however, are high sensitivity to the choice of handcrafted class names that define queries, and the difficulty of adaptation to new, smaller datasets. Towards addressing these problems, we propose to leverage available data to learn, for each class, an optimal word embedding as a function of the visual content. By learning new word embeddings on an otherwise frozen model, we are able to retain zero-shot capabilities for new classes, easily adapt models to new datasets, and adjust potentially erroneous, non-descriptive or ambiguous class names. We show that our solution can easily be integrated in image classification and object detection pipelines, yields significant performance gains in multiple scenarios and provides insights into model biases and labelling errors.

Self-supervised learning in vision--language processing (VLP) exploits semantic alignment between imaging and text modalities. Prior work in biomedical VLP has mostly relied on the alignment of single image and report pairs even though clinical notes commonly refer to prior images. This does not only introduce poor alignment between the modalities but also a missed opportunity to exploit rich self-supervision through existing temporal content in the data. In this work, we explicitly account for prior images and reports when available during both training and fine-tuning. Our approach, named BioViL-T, uses a CNN--Transformer hybrid multi-image encoder trained jointly with a text model. It is designed to be versatile to arising challenges such as pose variations and missing input images across time. The resulting model excels on downstream tasks both in single- and multi-image setups, achieving state-of-the-art (SOTA) performance on (I) progression classification, (II) phrase grounding, and (III) report generation, whilst offering consistent improvements on disease classification and sentence-similarity tasks. We release a novel multi-modal temporal benchmark dataset, CXR-T, to quantify the quality of vision--language representations in terms of temporal semantics. Our experimental results show the significant advantages of incorporating prior images and reports to make most use of the data.

The real-world data tends to be heavily imbalanced and severely skew the data-driven deep neural networks, which makes Long-Tailed Recognition (LTR) a massive challenging task. Existing LTR methods seldom train Vision Transformers (ViTs) with Long-Tailed (LT) data, while the off-the-shelf pretrain weight of ViTs always leads to unfair comparisons. In this paper, we systematically investigate the ViTs' performance in LTR and propose LiVT to train ViTs from scratch only with LT data. With the observation that ViTs suffer more severe LTR problems, we conduct Masked Generative Pretraining (MGP) to learn generalized features. With ample and solid evidence, we show that MGP is more robust than supervised manners. Although Binary Cross Entropy (BCE) loss performs well with ViTs, it struggles on the LTR tasks. We further propose the balanced BCE to ameliorate it with strong theoretical groundings. Specially, we derive the unbiased extension of Sigmoid and compensate extra logit margins for deploying it. Our Bal-BCE contributes to the quick convergence of ViTs in just a few epochs. Extensive experiments demonstrate that with MGP and Bal-BCE, LiVT successfully trains ViTs well without any additional data and outperforms comparable state-of-the-art methods significantly, e.g., our ViT-B achieves 81.0% Top-1 accuracy in iNaturalist 2018 without bells and whistles. Code is available at https://github.com/XuZhengzhuo/LiVT.