CVPR 2023 Papers

For augmented reality (AR), it is important that virtual assets appear to 'sit among' real world objects. The virtual element should variously occlude and be occluded by real matter, based on a plausible depth ordering. This occlusion should be consistent over time as the viewer's camera moves. Unfortunately, small mistakes in the estimated scene depth can ruin the downstream occlusion mask, and thereby the AR illusion. Especially in real-time settings, depths inferred near boundaries or across time can be inconsistent. In this paper, we challenge the need for depth-regression as an intermediate step. We instead propose an implicit model for depth and use that to predict the occlusion mask directly. The inputs to our network are one or more color images, plus the known depths of any virtual geometry. We show how our occlusion predictions are more accurate and more temporally stable than predictions derived from traditional depth-estimation models. We obtain state-of-the-art occlusion results on the challenging ScanNetv2 dataset and superior qualitative results on real scenes.

The last several years have witnessed remarkable progress in video-and-language (VidL) understanding. However, most modern VidL approaches use complex and specialized model architectures and sophisticated pretraining protocols, making the reproducibility, analysis and comparisons of these frameworks difficult. Hence, instead of proposing yet another new VidL model, this paper conducts a thorough empirical study demystifying the most important factors in the VidL model design. Among the factors that we investigate are (i) the spatiotemporal architecture design, (ii) the multimodal fusion schemes, (iii) the pretraining objectives, (iv) the choice of pretraining data, (v) pretraining and finetuning protocols, and (vi) dataset and model scaling. Our empirical study reveals that the most important design factors include: temporal modeling, video-to-text multimodal fusion, masked modeling objectives, and joint training on images and videos. Using these empirical insights, we then develop a step-by-step recipe, dubbed VindLU, for effective VidL pretraining. Our final model trained using our recipe achieves comparable or better than state-of-the-art results on several VidL tasks without relying on external CLIP pretraining. In particular, on the text-to-video retrieval task, our approach obtains 61.2% on DiDeMo, and 55.0% on ActivityNet, outperforming current SOTA by 7.8% and 6.1% respectively. Furthermore, our model also obtains state-of-the-art video question-answering results on ActivityNet-QA, MSRVTT-QA, MSRVTT-MC and TVQA. Our code and pretrained models are publicly available at: https://github.com/klauscc/VindLU.

3D object detection from visual sensors is a cornerstone capability of robotic systems. State-of-the-art methods focus on reasoning and decoding object bounding boxes from multi-view camera input. In this work we gain intuition from the integral role of multi-view consistency in 3D scene understanding and geometric learning. To this end, we introduce VEDet, a novel 3D object detection framework that exploits 3D multi-view geometry to improve localization through viewpoint awareness and equivariance. VEDet leverages a query-based transformer architecture and encodes the 3D scene by augmenting image features with positional encodings from their 3D perspective geometry. We design view-conditioned queries at the output level, which enables the generation of multiple virtual frames during training to learn viewpoint equivariance by enforcing multi-view consistency. The multi-view geometry injected at the input level as positional encodings and regularized at the loss level provides rich geometric cues for 3D object detection, leading to state-of-the-art performance on the nuScenes benchmark. The code and model are made available at https://github.com/TRI-ML/VEDet.

Although different approaches have been proposed for 3D point cloud-related tasks, few-shot learning (FSL) of 3D point clouds still remains under-explored. In FSL, unlike traditional supervised learning, the classes of training and test data do not overlap, and a model needs to recognize unseen classes from only a few samples. Existing FSL methods for 3D point clouds employ point-based models as their backbone. Yet, based on our extensive experiments and analysis, we first show that using a point-based backbone is not the most suitable FSL approach, since (i) a large number of points' features are discarded by the max pooling operation used in 3D point-based backbones, decreasing the ability of representing shape information; (ii)point-based backbones are sensitive to occlusion. To address these issues, we propose employing a projection- and 2D Convolutional Neural Network-based backbone, referred to as the ViewNet, for FSL from 3D point clouds. Our approach first projects a 3D point cloud onto six different views to alleviate the issue of missing points. Also, to generate more descriptive and distinguishing features, we propose View Pooling, which combines different projected plane combinations into five groups and performs max-pooling on each of them. The experiments performed on the ModelNet40, ScanObjectNN and ModelNet40-C datasets, with cross validation, show that our method consistently outperforms the state-of-the-art baselines. Moreover, compared to traditional image classification backbones, such as ResNet, the proposed ViewNet can extract more distinguishing features from multiple views of a point cloud. We also show that ViewNet can be used as a backbone with different FSL heads and provides improved performance compared to traditionally used backbones.

Existing Siamese tracking methods, which are built on pair-wise matching between two single frames, heavily rely on additional sophisticated mechanism to exploit temporal information among successive video frames, hindering them from high efficiency and industrial deployments. In this work, we resort to sequence-level target matching that can encode temporal contexts into the spatial features through a neat feedforward video model. Specifically, we adapt the standard video transformer architecture to visual tracking by enabling spatiotemporal feature learning directly from frame-level patch sequences. To better adapt to the tracking task, we carefully blend the spatiotemporal information in the video clips through sequential multi-branch triplet blocks, which formulates a video transformer backbone. Our experimental study compares different model variants, such as tokenization strategies, hierarchical structures, and video attention schemes. Then, we propose a disentangled dual-template mechanism that decouples static and dynamic appearance changes over time, and reduces the temporal redundancy in video frames. Extensive experiments show that our method, named as VideoTrack, achieves state-of-the-art results while running in real-time.

Scale is the primary factor for building a powerful foundation model that could well generalize to a variety of downstream tasks. However, it is still challenging to train video foundation models with billions of parameters. This paper shows that video masked autoencoder (VideoMAE) is a scalable and general self-supervised pre-trainer for building video foundation models. We scale the VideoMAE in both model and data with a core design. Specifically, we present a dual masking strategy for efficient pre-training, with an encoder operating on a subset of video tokens and a decoder processing another subset of video tokens. Although VideoMAE is very efficient due to high masking ratio in encoder, masking decoder can still further reduce the overall computational cost. This enables the efficient pre-training of billion-level models in video. We also introduce a progressive training paradigm that involves initial pre-training on the diverse multi-sourced unlabeled dataset, followed by fine-tuning on a mixed labeled dataset. Finally, we successfully train a video ViT model with a billion parameters, which achieves a new state-of-the-art performance on the datasets of Kinetics (90.0% on K400 and 89.9% on K600) and Something-Something (68.7% on V1 and 77.0% on V2). In addition, we extensively verify the pre-trained video ViT models on a variety of downstream tasks, demonstrating its effectiveness as a general video representation learner.

Contrastive learning-based video-language representation learning approaches, e.g., CLIP, have achieved outstanding performance, which pursue semantic interaction upon pre-defined video-text pairs. To clarify this coarse-grained global interaction and move a step further, we have to encounter challenging shell-breaking interactions for fine-grained cross-modal learning. In this paper, we creatively model video-text as game players with multivariate cooperative game theory to wisely handle the uncertainty during fine-grained semantic interaction with diverse granularity, flexible combination, and vague intensity. Concretely, we propose Hierarchical Banzhaf Interaction (HBI) to value possible correspondence between video frames and text words for sensitive and explainable cross-modal contrast. To efficiently realize the cooperative game of multiple video frames and multiple text words, the proposed method clusters the original video frames (text words) and computes the Banzhaf Interaction between the merged tokens. By stacking token merge modules, we achieve cooperative games at different semantic levels. Extensive experiments on commonly used text-video retrieval and video-question answering benchmarks with superior performances justify the efficacy of our HBI. More encouragingly, it can also serve as a visualization tool to promote the understanding of cross-modal interaction, which may have a far-reaching impact on the community. Project page is available at https://jpthu17.github.io/HBI/.

Although action recognition systems can achieve top performance when evaluated on in-distribution test points, they are vulnerable to unanticipated distribution shifts in test data. However, test-time adaptation of video action recognition models against common distribution shifts has so far not been demonstrated. We propose to address this problem with an approach tailored to spatio-temporal models that is capable of adaptation on a single video sample at a step. It consists in a feature distribution alignment technique that aligns online estimates of test set statistics towards the training statistics. We further enforce prediction consistency over temporally augmented views of the same test video sample. Evaluations on three benchmark action recognition datasets show that our proposed technique is architecture-agnostic and able to significantly boost the performance on both, the state of the art convolutional architecture TANet and the Video Swin Transformer. Our proposed method demonstrates a substantial performance gain over existing test-time adaptation approaches in both evaluations of a single distribution shift and the challenging case of random distribution shifts.

Despite the remarkable progress in deep generative models, synthesizing high-resolution and temporally coherent videos still remains a challenge due to their high-dimensionality and complex temporal dynamics along with large spatial variations. Recent works on diffusion models have shown their potential to solve this challenge, yet they suffer from severe computation- and memory-inefficiency that limit the scalability. To handle this issue, we propose a novel generative model for videos, coined projected latent video diffusion models (PVDM), a probabilistic diffusion model which learns a video distribution in a low-dimensional latent space and thus can be efficiently trained with high-resolution videos under limited resources. Specifically, PVDM is composed of two components: (a) an autoencoder that projects a given video as 2D-shaped latent vectors that factorize the complex cubic structure of video pixels and (b) a diffusion model architecture specialized for our new factorized latent space and the training/sampling procedure to synthesize videos of arbitrary length with a single model. Experiments on popular video generation datasets demonstrate the superiority of PVDM compared with previous video synthesis methods; e.g., PVDM obtains the FVD score of 639.7 on the UCF-101 long video (128 frames) generation benchmark, which improves 1773.4 of the prior state-of-the-art.

Video anomaly detection (VAD) is a significant computer vision problem. Existing deep neural network (DNN) based VAD methods mostly follow the route of frame reconstruction or frame prediction. However, the lack of mining and learning of higher-level visual features and temporal context relationships in videos limits the further performance of these two approaches. Inspired by video codec theory, we introduce a brand-new VAD paradigm to break through these limitations: First, we propose a new task of video event restoration based on keyframes. Encouraging DNN to infer missing multiple frames based on video keyframes so as to restore a video event, which can more effectively motivate DNN to mine and learn potential higher-level visual features and comprehensive temporal context relationships in the video. To this end, we propose a novel U-shaped Swin Transformer Network with Dual Skip Connections (USTN-DSC) for video event restoration, where a cross-attention and a temporal upsampling residual skip connection are introduced to further assist in restoring complex static and dynamic motion object features in the video. In addition, we propose a simple and effective adjacent frame difference loss to constrain the motion consistency of the video sequence. Extensive experiments on benchmarks demonstrate that USTN-DSC outperforms most existing methods, validating the effectiveness of our method.

Video dehazing aims to recover haze-free frames with high visibility and contrast. This paper presents a novel framework to effectively explore the physical haze priors and aggregate temporal information. Specifically, we design a memory-based physical prior guidance module to encode the prior-related features into long-range memory. Besides, we formulate a multi-range scene radiance recovery module to capture space-time dependencies in multiple space-time ranges, which helps to effectively aggregate temporal information from adjacent frames. Moreover, we construct the first large-scale outdoor video dehazing benchmark dataset, which contains videos in various real-world scenarios. Experimental results on both synthetic and real conditions show the superiority of our proposed method.

Encoding videos as neural networks is a recently proposed approach that allows new forms of video processing. However, traditional techniques still outperform such neural video representation (NVR) methods for the task of video compression. This performance gap can be explained by the fact that current NVR methods: i) use architectures that do not efficiently obtain a compact representation of temporal and spatial information; and ii) minimize rate and distortion disjointly (first overfitting a network on a video and then using heuristic techniques such as post-training quantization or weight pruning to compress the model). We propose a novel convolutional architecture for video representation that better represents spatio-temporal information and a training strategy capable of jointly optimizing rate and distortion. All network and quantization parameters are jointly learned end-to-end, and the post-training operations used in previous works are unnecessary. We evaluate our method on the UVG dataset, achieving new state-of-the-art results for video compression with NVRs. Moreover, we deliver the first NVR-based video compression method that improves over the typically adopted HEVC benchmark (x265, disabled b-frames, "medium" preset), closing the gap to autoencoder-based video compression techniques.

In this work, we introduce Vid2Seq, a multi-modal single-stage dense event captioning model pretrained on narrated videos which are readily-available at scale. The Vid2Seq architecture augments a language model with special time tokens, allowing it to seamlessly predict event boundaries and textual descriptions in the same output sequence. Such a unified model requires large-scale training data, which is not available in current annotated datasets. We show that it is possible to leverage unlabeled narrated videos for dense video captioning, by reformulating sentence boundaries of transcribed speech as pseudo event boundaries, and using the transcribed speech sentences as pseudo event captions. The resulting Vid2Seq model pretrained on the YT-Temporal-1B dataset improves the state of the art on a variety of dense video captioning benchmarks including YouCook2, ViTT and ActivityNet Captions. Vid2Seq also generalizes well to the tasks of video paragraph captioning and video clip captioning, and to few-shot settings. Our code is publicly available at https://antoyang.github.io/vid2seq.html.

We present Vid2Avatar, a method to learn human avatars from monocular in-the-wild videos. Reconstructing humans that move naturally from monocular in-the-wild videos is difficult. Solving it requires accurately separating humans from arbitrary backgrounds. Moreover, it requires reconstructing detailed 3D surface from short video sequences, making it even more challenging. Despite these challenges, our method does not require any groundtruth supervision or priors extracted from large datasets of clothed human scans, nor do we rely on any external segmentation modules. Instead, it solves the tasks of scene decomposition and surface reconstruction directly in 3D by modeling both the human and the background in the scene jointly, parameterized via two separate neural fields. Specifically, we define a temporally consistent human representation in canonical space and formulate a global optimization over the background model, the canonical human shape and texture, and per-frame human pose parameters. A coarse-to-fine sampling strategy for volume rendering and novel objectives are introduced for a clean separation of dynamic human and static background, yielding detailed and robust 3D human reconstructions. The evaluation of our method shows improvements over prior art on publicly available datasets.

In this paper we introduce the Temporo-Spatial Vision Transformer (TSViT), a fully-attentional model for general Satellite Image Time Series (SITS) processing based on the Vision Transformer (ViT). TSViT splits a SITS record into non-overlapping patches in space and time which are tokenized and subsequently processed by a factorized temporo-spatial encoder. We argue, that in contrast to natural images, a temporal-then-spatial factorization is more intuitive for SITS processing and present experimental evidence for this claim. Additionally, we enhance the model's discriminative power by introducing two novel mechanisms for acquisition-time-specific temporal positional encodings and multiple learnable class tokens. The effect of all novel design choices is evaluated through an extensive ablation study. Our proposed architecture achieves state-of-the-art performance, surpassing previous approaches by a significant margin in three publicly available SITS semantic segmentation and classification datasets. All model, training and evaluation codes can be found at https://github.com/michaeltrs/DeepSatModels.

Human-Object Interaction (HOI) detection, which localizes and infers relationships between human and objects, plays an important role in scene understanding. Although two-stage HOI detectors have advantages of high efficiency in training and inference, they suffer from lower performance than one-stage methods due to the old backbone networks and the lack of considerations for the HOI perception process of humans in the interaction classifiers. In this paper, we propose Vision Transformer based Pose-Conditioned Self-Loop Graph (ViPLO) to resolve these problems. First, we propose a novel feature extraction method suitable for the Vision Transformer backbone, called masking with overlapped area (MOA) module. The MOA module utilizes the overlapped area between each patch and the given region in the attention function, which addresses the quantization problem when using the Vision Transformer backbone. In addition, we design a graph with a pose-conditioned self-loop structure, which updates the human node encoding with local features of human joints. This allows the classifier to focus on specific human joints to effectively identify the type of interaction, which is motivated by the human perception process for HOI. As a result, ViPLO achieves the state-of-the-art results on two public benchmarks, especially obtaining a +2.07 mAP performance gain on the HICO-DET dataset.

Perception and prediction are two separate modules in the existing autonomous driving systems. They interact with each other via hand-picked features such as agent bounding boxes and trajectories. Due to this separation, prediction, as a downstream module, only receives limited information from the perception module. To make matters worse, errors from the perception modules can propagate and accumulate, adversely affecting the prediction results. In this work, we propose ViP3D, a query-based visual trajectory prediction pipeline that exploits rich information from raw videos to directly predict future trajectories of agents in a scene. ViP3D employs sparse agent queries to detect, track, and predict throughout the pipeline, making it the first fully differentiable vision-based trajectory prediction approach. Instead of using historical feature maps and trajectories, useful information from previous timestamps is encoded in agent queries, which makes ViP3D a concise streaming prediction method. Furthermore, extensive experimental results on the nuScenes dataset show the strong vision-based prediction performance of ViP3D over traditional pipelines and previous end-to-end models.

Dominant pre-training works for image-text retrieval adopt "dual-encoder" architecture to enable high efficiency, where two encoders are used to extract image and text representations and contrastive learning is employed for global alignment. However, coarse-grained global alignment ignores detailed semantic associations between image and text. In this work, we propose a novel proxy task, named Visual-Language Error Modeling (ViLEM), to inject detailed image-text association into "dual-encoder" model by "proofreading" each word in the text against the corresponding image. Specifically, we first edit the image-paired text to automatically generate diverse plausible negative texts with pre-trained language models. ViLEM then enforces the model to discriminate the correctness of each word in the plausible negative texts and further correct the wrong words via resorting to image information. Furthermore, we propose a multi-granularity interaction framework to perform ViLEM via interacting text features with both global and local image features, which associates local text semantics with both high-level visual context and multi-level local visual information. Our method surpasses state-of-the-art "dual-encoder" methods by a large margin on the image-text retrieval task and significantly improves discriminativeness to local textual semantics. Our model can also generalize well to video-text retrieval.