CVPR 2024 Papers

Generating novel views of an object from a single image is a challenging task. It requires an understanding of the underlying 3D structure of the object from an image and rendering high-quality spatially consistent new views. While recent methods for view synthesis based on diffusion have shown great progress achieving consistency among various view estimates and at the same time abiding by the desired camera pose remains a critical problem yet to be solved. In this work we demonstrate a strikingly simple method where we utilize a pre-trained video diffusion model to solve this problem. Our key idea is that synthesizing a novel view could be reformulated as synthesizing a video of a camera going around the object of interest---a scanning video---which then allows us to leverage the powerful priors that a video diffusion model would have learned. Thus to perform novel-view synthesis we create a smooth camera trajectory to the target view that we wish to render and denoise using both a view-conditioned diffusion model and a video diffusion model. By doing so we obtain a highly consistent novel view synthesis outperforming the state of the art.

The estimation of implicit cross-frame correspondences and the high computational cost have long been major challenges in video semantic segmentation (VSS) for driving scenes. Prior works utilize keyframes feature propagation or cross-frame attention to address these issues. By contrast we are the first to harness vanishing point (VP) priors for more effective segmentation. Intuitively objects near VPs (i.e. away from the vehicle) are less discernible. Moreover they tend to move radially away from the VP over time in the usual case of a forward-facing camera a straight road and linear forward motion of the vehicle. Our novel efficient network for VSS named VPSeg incorporates two modules that utilize exactly this pair of static and dynamic VP priors: sparse-to-dense feature mining (DenseVP) and VP-guided motion fusion (MotionVP). MotionVP employs VP-guided motion estimation to establish explicit correspondences across frames and help attend to the most relevant features from neighboring frames while DenseVP enhances weak dynamic features in distant regions around VPs. These modules operate within a context-detail framework which separates contextual features from high-resolution local features at different input resolutions to reduce computational costs. Contextual and local features are integrated through contextualized motion attention (CMA) for the final prediction. Extensive experiments on two popular driving segmentation benchmarks Cityscapes and ACDC demonstrate that VPSeg outperforms previous SOTA methods with only modest computational overhead.

Large language models (LLMs) have shown remarkable text understanding capabilities which have been extended as Video LLMs to handle video data for comprehending visual details. However existing Video LLMs can only provide a coarse description of the entire video failing to capture the precise start and end time boundary of specific events. In this paper we solve this issue via proposing VTimeLLM a novel Video LLM designed for fine-grained video moment understanding and reasoning with respect to time boundary. Specifically our VTimeLLM adopts a boundary-aware three-stage training strategy which respectively utilizes image-text pairs for feature alignment multiple-event videos to increase temporal-boundary awareness and high-quality video-instruction tuning to further improve temporal understanding ability as well as align with human intents. Extensive experiments demonstrate that in fine-grained time-related comprehension tasks for videos such as Temporal Video Grounding and Dense Video Captioning VTimeLLM significantly outperforms existing Video LLMs. Besides benefits from the fine-grained temporal understanding of the videos further enable VTimeLLM to beat existing Video LLMs in video dialogue benchmark showing its superior cross-modal understanding and reasoning abilities.

Text-to-video diffusion models have advanced video generation significantly. However customizing these models to generate videos with tailored motions presents a substantial challenge. In specific they encounter hurdles in (1) accurately reproducing motion from a target video and (2) creating diverse visual variations. For example straightforward extensions of static image customization methods to video often lead to intricate entanglements of appearance and motion data. To tackle this here we present the Video Motion Customization (VMC) framework a novel one-shot tuning approach crafted to adapt temporal attention layers within video diffusion models. Our approach introduces a novel motion distillation objective using residual vectors between consecutive noisy latent frames as a motion reference. The diffusion process then preserve low-frequency motion trajectories while mitigating high-frequency motion-unrelated noise in image space. We validate our method against state-of-the-art video generative models across diverse real-world motions and contexts. Our codes and data can be found at: https://video-motion-customization.github.io/

Rigging and skinning clothed human avatars is a challenging task and traditionally requires a lot of manual work and expertise. Recent methods addressing it either generalize across different characters or focus on capturing the dynamics of a single character observed under different pose configurations. However the former methods typically predict solely static skinning weights which perform poorly for highly articulated poses and the latter ones either require dense 3D character scans in different poses or cannot generate an explicit mesh with vertex correspondence over time. To address these challenges we propose a fully automated approach for creating a fully rigged character with pose-dependent skinning weights which can be solely learned from multi-view video. Therefore we first acquire a rigged template which is then statically skinned. Next a coordinate-based MLP learns a skinning weights field parameterized over the position in a canonical pose space and the respective pose. Moreover we introduce our pose- and view-dependent appearance field allowing us to differentiably render and supervise the posed mesh using multi-view imagery. We show that our approach outperforms state-of-the-art while not relying on dense 4D scans. More details can be found on our project page.

Video generation has witnessed significant advancements yet evaluating these models remains a challenge. A comprehensive evaluation benchmark for video generation is indispensable for two reasons: 1) Existing metrics do not fully align with human perceptions; 2) An ideal evaluation system should provide insights to inform future developments of video generation. To this end we present VBench a comprehensive benchmark suite that dissects "video generation quality" into specific hierarchical and disentangled dimensions each with tailored prompts and evaluation methods. VBench has three appealing properties: 1) Comprehensive Dimensions: VBench comprises 16 dimensions in video generation (e.g. subject identity inconsistency motion smoothness temporal flickering and spatial relationship etc). The evaluation metrics with fine-grained levels reveal individual models' strengths and weaknesses. 2) Human Alignment: We also provide a dataset of human preference annotations to validate our benchmarks' alignment with human perception for each evaluation dimension respectively. 3) Valuable Insights: We look into current models' ability across various evaluation dimensions and various content types. We also investigate the gaps between video and image generation models. We will open-source VBench including all prompts evaluation methods generated videos and human preference annotations and also include more video generation models in VBench to drive forward the field of video generation.

Text-based diffusion models have exhibited remarkable success in generation and editing showing great promise for enhancing visual content with their generative prior. However applying these models to video super-resolution remains challenging due to the high demands for output fidelity and temporal consistency which is complicated by the inherent randomness in diffusion models. Our study introduces Upscale-A-Video a text-guided latent diffusion framework for video upscaling. This framework ensures temporal coherence through two key mechanisms: locally it integrates temporal layers into U-Net and VAE-Decoder maintaining consistency within short sequences; globally without training a flow-guided recurrent latent propagation module is introduced to enhance overall video stability by propagating and fusing latent across the entire sequences. Thanks to the diffusion paradigm our model also offers greater flexibility by allowing text prompts to guide texture creation and adjustable noise levels to balance restoration and generation enabling a trade-off between fidelity and quality. Extensive experiments show that Upscale-A-Video surpasses existing methods in both synthetic and real-world benchmarks as well as in AI-generated videos showcasing impressive visual realism and temporal consistency.

In this work we tackle the problem of unsupervised domain adaptation (UDA) for video action recognition. Our approach which we call UNITE uses an image teacher model to adapt a video student model to the target domain. UNITE first employs self-supervised pre-training to promote discriminative feature learning on target domain videos using a teacher-guided masked distillation objective. We then perform self-training on masked target data using the video student model and image teacher model together to generate improved pseudolabels for unlabeled target videos. Our self-training process successfully leverages the strengths of both models to achieve strong transfer performance across domains. We evaluate our approach on multiple video domain adaptation benchmarks and observe significant improvements upon previously reported results.

Despite the recent advances in unified image segmentation (IS) developing a unified video segmentation (VS) model remains a challenge. This is mainly because generic category-specified VS tasks need to detect all objects and track them across consecutive frames while prompt-guided VS tasks require re-identifying the target with visual/text prompts throughout the entire video making it hard to handle the different tasks with the same architecture. We make an attempt to address these issues and present a novel unified VS architecture namely UniVS by using prompts as queries. UniVS averages the prompt features of the target from previous frames as its initial query to explicitly decode masks and introduces a target-wise prompt cross-attention layer in the mask decoder to integrate prompt features in the memory pool. By taking the predicted masks of entities from previous frames as their visual prompts UniVS converts different VS tasks into prompt-guided target segmentation eliminating the heuristic inter-frame matching process. Our framework not only unifies the different VS tasks but also naturally achieves universal training and testing ensuring robust performance across different scenarios. UniVS shows a commendable balance between performance and universality on 10 challenging VS benchmarks covering video instance semantic panoptic object and referring segmentation tasks. Code can be found at https://github.com/MinghanLi/UniVS.

Source-free Domain Adaptation (SFDA) is an emerging and challenging research area that addresses the problem of unsupervised domain adaptation (UDA) without source data. Though numerous successful methods have been proposed for SFDA a theoretical understanding of why these methods work well is still absent. In this paper we shed light on the theoretical perspective of existing SFDA methods. Specifically we find that SFDA loss functions comprising discriminability and diversity losses work in the same way as the training objective in the theory of self-training based on the expansion assumption which shows the existence of the target error bound. This finding brings two novel insights that enable us to build an improved SFDA method comprising 1) Model Training with Auto-Adjusting Diversity Constraint and 2) Augmentation Training with Teacher-Student Framework yielding a better recognition performance. Extensive experiments on three benchmark datasets demonstrate the validity of the theoretical analysis and our method.

This paper studies the problem of concept-based interpretability of transformer representations for videos. Concretely we seek to explain the decision-making process of video transformers based on high-level spatiotemporal concepts that are automatically discovered. Prior research on concept-based interpretability has concentrated solely on image-level tasks. Comparatively video models deal with the added temporal dimension increasing complexity and posing challenges in identifying dynamic concepts over time. In this work we systematically address these challenges by introducing the first Video Transformer Concept Discovery (VTCD) algorithm. To this end we propose an efficient approach for unsupervised identification of units of video transformer representations - concepts and ranking their importance to the output of a model. The resulting concepts are highly interpretable revealing spatio-temporal reasoning mechanisms and object-centric representations in unstructured video models. Performing this analysis jointly over a diverse set of supervised and self-supervised representations we discover that some of these mechanism are universal in video transformers. Finally we show that VTCD can be used for fine-grained action recognition and video object segmentation.

Learning-based underwater image enhancement (UIE) methods have made great progress. However the lack of large-scale and high-quality paired training samples has become the main bottleneck hindering the development of UIE. The inter-frame information in underwater videos can accelerate or optimize the UIE process. Thus we constructed the first large-scale high-resolution underwater video enhancement benchmark (UVEB) to promote the development of underwater vision.It contains 1308 pairs of video sequences and more than 453000 high-resolution with 38% Ultra-High-Definition (UHD) 4K frame pairs. UVEB comes from multiple countries containing various scenes and video degradation types to adapt to diverse and complex underwater environments. We also propose the first supervised underwater video enhancement method UVE-Net. UVE-Net converts the current frame information into convolutional kernels and passes them to adjacent frames for efficient inter-frame information exchange. By fully utilizing the redundant degraded information of underwater videos UVE-Net completes video enhancement better. Experiments show the effective network design and good performance of UVE-Net.

Recent advancements in multimodal pre-training have shown promising efficacy in 3D representation learning by aligning multimodal features across 3D shapes their 2D counterparts and language descriptions. However the methods used by existing frameworks to curate such multimodal data in particular language descriptions for 3D shapes are not scalable and the collected language descriptions are not diverse. To address this we introduce ULIP-2 a simple yet effective tri-modal pretraining framework that leverages large multimodal models to automatically generate holistic language descriptions for 3D shapes. It only needs 3D data as input eliminating the need for any manual 3D annotations and is therefore scalable to large datasets. ULIP-2 is also equipped with scaled-up backbones for better multi-modal representation learning. We conduct experiments on two large-scale 3D datasets Objaverse and ShapeNet and augment them with tri-modal datasets of 3D point clouds images and language for training ULIP-2. Experiments show that ULIP-2 demonstrates substantial benefits in three downstream tasks: zero-shot 3D classification standard 3D classification with fine-tuning and 3D captioning (3D-to-language generation). It achieves a new SOTA of 50.6% (top- 1) on Objaverse-LVIS and 84.7% (top-1) on ModelNet40 in zero-shot classification. In the ScanObjectNN benchmark for standard fine-tuning ULIP-2 reaches an overall accuracy of 91.5% with a compact model of only 1.4 million parameters. ULIP-2 sheds light on a new paradigm for scalable multimodal 3D representation learning without human annotations and shows significant improvements over existing baselines. The code and datasets are released at https://github.com/salesforce/ULIP.

Recent literature has demonstrated that vision transformers (VITs) exhibit superior performance compared to convolutional neural networks (CNNs). The majority of recent research on adversarial robustness however has predominantly focused on CNNs. In this work we bridge this gap by analyzing the effectiveness of existing attacks on VITs. We demonstrate that due to the softmax computations in every attention block in VITs they are inherently vulnerable to floating point underflow errors. This can lead to a gradient masking effect resulting in suboptimal attack strength of well-known attacks like PGD Carlini and Wagner (CW) GAMA and Patch attacks. Motivated by this we propose Adaptive Attention Scaling (AAS) attack that can automatically find the optimal scaling factors of pre-softmax outputs using gradient-based optimization. We show that the proposed simple strategy can be incorporated with any existing adversarial attacks as well as adversarial training methods and achieved improved performance. On VIT-B16 we demonstrate an improved attack strength of upto 2.2% on CIFAR10 and upto 2.9% on CIFAR100 by incorporating the proposed AAS attack with state-of-the-art single attack methods like GAMA attack. Further we utilise the proposed AAS attack for every few epochs in existing adversarial training methods which is termed as Adaptive Attention Scaling Adversarial Training (AAS-AT). On incorporating AAS-AT with existing methods we outperform them on VITs over 1.3-3.5% on CIFAR10. We observe improved performance on ImageNet-100 as well.

Surveillance videos are important for public security. However current surveillance video tasks mainly focus on classifying and localizing anomalous events. Existing methods are limited to detecting and classifying the predefined events with unsatisfactory semantic understanding although they have obtained considerable performance. To address this issue we propose a new research direction of surveillance video-and-language understanding(VALU) and construct the first multimodal surveillance video dataset. We manually annotate the real-world surveillance dataset UCF-Crime with fine-grained event content and timing. Our newly annotated dataset UCA (UCF-Crime Annotation) contains 23542 sentences with an average length of 20 words and its annotated videos are as long as 110.7 hours. Furthermore we benchmark SOTA models for four multimodal tasks on this newly created dataset which serve as new baselines for surveillance VALU. Through experiments we find that mainstream models used in previously public datasets perform poorly on surveillance video demonstrating new challenges in surveillance VALU. We also conducted experiments on multimodal anomaly detection. These results demonstrate that our multimodal surveillance learning can improve the performance of anomaly detection. All the experiments highlight the necessity of constructing this dataset to advance surveillance AI.

The rapid advancement of large language models (LLMs) has accelerated the emergence of in-context learning (ICL) as a cutting-edge approach in the natural language processing domain. Recently ICL has been employed in visual understanding tasks such as semantic segmentation and image captioning yielding promising results. However existing visual ICL framework can not enable producing content across multiple modalities which limits their potential usage scenarios. To address this issue we present a new ICL framework for visual understanding with multi-modal output enabled. First we quantize and embed both text and visual prompt into a unified representational space structured as interleaved in-context sequences. Then a decoder-only sparse transformer architecture is employed to perform generative modeling on them facilitating in-context learning. Thanks to this design the model is capable of handling in-context vision understanding tasks with multimodal output in a unified pipeline. Experimental results demonstrate that our model achieves competitive performance compared with specialized models and previous ICL baselines. Overall our research takes a further step toward unified multimodal in-context learning.

We introduce a new task -- language-driven video inpainting which uses natural language instructions to guide the inpainting process. This approach overcomes the limitations of traditional video inpainting methods that depend on manually labeled binary masks a process often tedious and labor-intensive. We present the Remove Objects from Videos by Instructions (ROVI) dataset containing 5650 videos and 9091 inpainting results to support training and evaluation for this task. We also propose a novel diffusion-based language-driven video inpainting framework the first end-to-end baseline for this task integrating Multimodal Large Language Models to understand and execute complex language-based inpainting requests effectively. Our comprehensive results showcase the dataset's versatility and the model's effectiveness in various language-instructed inpainting scenarios. We have made datasets code and models publicly available at https://github.com/jianzongwu/Language-Driven-Video-Inpainting.

The spiking cameras offer the benefits of high dynamic range (HDR) high temporal resolution and low data redundancy. However reconstructing HDR videos in high-speed conditions using single-bit spikings presents challenges due to the limited bit depth. Increasing the bit depth of the spikings is advantageous for boosting HDR performance but the readout efficiency will be decreased which is unfavorable for achieving a high frame rate (HFR) video. To address these challenges we propose a readout mechanism to obtain rolling-mixed-bit (RMB) spikings which involves interleaving multi-bit spikings within the single-bit spikings in a rolling manner thereby combining the characteristics of high bit depth and efficient readout. Furthermore we introduce RMB-Net for reconstructing HDR and HFR videos. RMB-Net comprises a cross-bit attention block for fusing mixed-bit spikings and a cross-time attention block for achieving temporal fusion. Extensive experiments conducted on synthetic and real-synthetic data demonstrate the superiority of our method. For instance pure 3-bit spikings result in 3 times of data volume whereas our method achieves comparable performance with less than 2% increase in data volume.