CVPR 2024 论文列表

With the rapid development of Multi-modal Large Language Models (MLLMs) a number of diagnostic benchmarks have recently emerged to evaluate the comprehension capabilities of these models. However most benchmarks predominantly assess spatial understanding in the static image tasks while overlooking temporal understanding in the dynamic video tasks. To alleviate this issue we introduce a comprehensive Multi-modal Video understanding Benchmark namely MVBench which covers 20 challenging video tasks that cannot be effectively solved with a single frame. Specifically we first introduce a novel static-to-dynamic method to define these temporal-related tasks. By transforming various static tasks into dynamic ones we enable the systematic generation of video tasks that require a broad spectrum of temporal skills ranging from perception to cognition. Then guided by the task definition we automatically convert public video annotations into multiple-choice QA to evaluate each task. On one hand such a distinct paradigm allows us to build MVBench efficiently without much manual intervention. On the other hand it guarantees evaluation fairness with ground-truth video annotations avoiding the biased scoring of LLMs. Moreover we further develop a robust video MLLM baseline i.e. VideoChat2 by progressive multi-modal training with diverse instruction-tuning data. The extensive results on our MVBench reveal that the existing MLLMs are far from satisfactory in temporal understanding while our VideoChat2 largely surpasses these leading models by over 15% on MVBench.

State-of-the-art video-text retrieval (VTR) methods typically involve fully fine-tuning a pre-trained model (e.g. CLIP) on specific datasets. However this can result in significant storage costs in practical applications as a separate model per task must be stored. To address this issue we present our pioneering work that enables parameter-efficient VTR using a pre-trained model with only a small number of tunable parameters during training. Towards this goal we propose a new method dubbed Multimodal Video Adapter (MV-Adapter) for efficiently transferring the knowledge in the pre-trained CLIP from image-text to video-text. Specifically MV-Adapter utilizes bottleneck structures in both video and text branches along with two novel components. The first is a Temporal Adaptation Module that is incorporated in the video branch to introduce global and local temporal contexts. We also train weights calibrations to adjust to dynamic variations across frames. The second is Cross Modality Tying that generates weights for video/text branches through sharing cross modality factors for better aligning between modalities. Thanks to above innovations MV-Adapter can achieve comparable or better performance than standard fine-tuning with negligible parameters overhead. Notably MV-Adapter consistently outperforms various competing methods in V2T/T2V tasks with large margins on five widely used VTR benchmarks (MSR-VTT MSVD LSMDC DiDemo and ActivityNet). Codes will be released.

We introduce MMMU: a new benchmark designed to evaluate multimodal models on massive multi-discipline tasks demanding college-level subject knowledge and deliberate reasoning. MMMU includes 11.5K meticulously collected multimodal questions from college exams quizzes and textbooks covering six core disciplines: Art & Design Business Science Health & Medicine Humanities & Social Science and Tech & Engineering. These questions span 30 subjects and 183 subfields comprising 30 highly heterogeneous image types such as charts diagrams maps tables music sheets and chemical structures. Unlike existing benchmarks MMMU focuses on advanced perception and reasoning with domain-specific knowledge challenging models to perform tasks akin to those faced by experts. The evaluation of 28 open-source LMMs as well as the proprietary GPT-4V(ision) and Gemini highlights the substantial challenges posed by MMMU. Even the advanced GPT-4V and Gemini Ultra only achieve accuracies of 56% and 59% respectively indicating significant room for improvement. We believe MMMU will stimulate the community to build next-generation multimodal foundation models towards expert artificial general intelligence.

With the success of large language models (LLMs) integrating the vision model into LLMs to build vision-language foundation models has gained much more interest recently. However existing LLM-based large multimodal models (e.g. Video-LLaMA VideoChat) can only take in a limited number of frames for short video understanding. In this study we mainly focus on designing an efficient and effective model for long-term video understanding. Instead of trying to process more frames simultaneously like most existing work we propose to process videos in an online manner and store past video information in a memory bank. This allows our model to reference historical video content for long-term analysis without exceeding LLMs' context length constraints or GPU memory limits. Our memory bank can be seamlessly integrated into current multimodal LLMs in an off-the-shelf manner. We conduct extensive experiments on various video understanding tasks such as long-video understanding video question answering and video captioning and our model can achieve state-of-the-art performances across multiple datasets.

Unprocessed RAW video has shown distinct advantages over sRGB video in video editing and computer vision tasks. However capturing RAW video is challenging due to limitations in bandwidth and storage. Various methods have been proposed to address similar issues in single image RAW capture through de-rendering. These methods utilize both the metadata and the sRGB image to perform sRGB-to-RAW de-rendering and recover high-quality single-frame RAW data. However metadata-based methods always require additional computation for online metadata generation imposing severe burden on mobile camera device for high frame rate RAW video capture. To address this issue we propose a framework that utilizes frame affinity to achieve high-quality sRGB-to-RAW video reconstruction. Our approach consists of two main steps. The first step temporal affinity prior extraction uses motion information between adjacent frames to obtain a reference RAW image. The second step spatial feature fusion and mapping learns a pixel-level mapping function using scene-specific and position-specific features provided by the previous frame. Our method can be easily applied to current mobile camera equipment without complicated adaptations or added burden. To demonstrate the effectiveness of our approach we introduce the first RAW Video De-rendering Benchmark. In this benchmark our method outperforms state-of-the-art RAW image reconstruction methods even without image-level metadata.

Multiple camera view (multi-view) setups have proven useful in many computer vision applications. However the high computational cost associated with multiple views creates a significant challenge for end devices with limited computational resources. In modern CPU pipelining breaks a longer job into steps and enables parallelism over sequential steps from multiple jobs. Inspired by this we study selective view pipelining for efficient multi-view understanding which breaks computation of multiple views into steps and only computes the most helpful views/steps in a parallel manner for the best efficiency. To this end we use reinforcement learning to learn a very light view selection module that analyzes the target object or scenario from initial views and selects the next-best-view for recognition or detection for pipeline computation. Experimental results on multi-view classification and detection tasks show that our approach achieves promising performance while using only 2 or 3 out of N available views significantly reducing computational costs while maintaining parallelism over GPU through selective view pipelining.

In this paper we tackle the problem of self-supervised video alignment and activity progress prediction using in-the-wild videos. Our proposed self-supervised representation learning method carefully addresses different action orderings redundant actions and background frames to generate improved video representations compared to previous methods. Our model generalizes temporal cycle-consistency learning to allow for more flexibility in determining cycle-consistent neighbors. More specifically to handle repeated actions we propose a multi-neighbor cycle consistency and a multi-cycle-back regression loss by finding multiple soft nearest neighbors using a Gaussian Mixture Model. To handle background and redundant frames we introduce a context-dependent drop function in our framework discouraging the alignment of droppable frames. On the other hand to learn from videos of multiple activities jointly we propose a multi-head crosstask network allowing us to embed a video and estimate progress without knowing its activity label. Experiments on multiple datasets show that our method outperforms the state-of-the-art for video alignment and progress prediction.

We introduce a framework for online learning from a single continuous video stream - the way people and animals learn without mini-batches data augmentation or shuffling. This poses great challenges given the high correlation between consecutive video frames and there is very little prior work on it. Our framework allows us to do a first deep dive into the topic and includes a collection of streams and tasks composed from two existing video datasets plus methodology for performance evaluation that considers both adaptation and generalization. We employ pixel-to-pixel modelling as a practical and flexible way to switch between pre-training and single-stream evaluation as well as between arbitrary tasks without ever requiring changes to models and always using the same pixel loss. Equipped with this framework we obtained large single-stream learning gains from pre-training with a novel family of future prediction tasks found that momentum hurts and that the pace of weight updates matters. The combination of these insights leads to matching the performance of IID learning with batch size 1 when using the same architecture and without costly replay buffers.

Diffusion models are just at a tipping point for image super-resolution task. Nevertheless it is not trivial to capitalize on diffusion models for video super-resolution which necessitates not only the preservation of visual appearance from low-resolution to high-resolution videos but also the temporal consistency across video frames. In this paper we propose a novel approach pursuing Spatial Adaptation and Temporal Coherence (SATeCo) for video super-resolution. SATeCo pivots on learning spatial-temporal guidance from low-resolution videos to calibrate both latent-space high-resolution video denoising and pixel-space video reconstruction. Technically SATeCo freezes all the parameters of the pre-trained UNet and VAE and only optimizes two deliberately-designed spatial feature adaptation (SFA) and temporal feature alignment (TFA) modules in the decoder of UNet and VAE. SFA modulates frame features via adaptively estimating affine parameters for each pixel guaranteeing pixel-wise guidance for high-resolution frame synthesis. TFA delves into feature interaction within a 3D local window (tubelet) through self-attention and executes cross-attention between tubelet and its low-resolution counterpart to guide temporal feature alignment. Extensive experiments conducted on the REDS4 and Vid4 datasets demonstrate the effectiveness of our approach.

Panoramic videos have the advantage of providing an immersive and interactive viewing experience. Nevertheless their spherical nature gives rise to various and uncertain user viewing behaviors which poses significant challenges for panoramic video quality assessment (PVQA). In this work we propose an end-to-end optimized blind PVQA method with explicit modeling of user viewing patterns through visual scanpaths. Our method consists of two modules: a scanpath generator and a quality assessor. The scanpath generator is initially trained to predict future scanpaths by minimizing their expected code length and then jointly optimized with the quality assessor for quality prediction. Our blind PVQA method enables direct quality assessment of panoramic images by treating them as videos composed of identical frames. Experiments on three public panoramic image and video quality datasets encompassing both synthetic and authentic distortions validate the superiority of our blind PVQA model over existing methods.

Existing scene text detectors generally focus on accurately detecting single-level (i.e. word-level line-level or paragraph-level) text entities without exploring the relationships among different levels of text entities. To comprehensively understand scene texts detecting multi-level texts while exploring their contextual information is critical. To this end we propose a unified framework (dubbed LayoutFormer) for hierarchical text detection which simultaneously conducts multi-level text detection and predicts the geometric layouts for promoting scene text understanding. In LayoutFormer WordDecoder LineDecoder and ParaDecoder are proposed to be responsible for word-level text prediction line-level text prediction and paragraph-level text prediction respectively. Meanwhile WordDecoder and ParaDecoder adaptively learn word-line and line-paragraph relationships respectively. In addition we propose a Prior Location Sampler to be used on multi-scale features to adaptively select a few representative foreground features for updating text queries. It can improve hierarchical detection performance while significantly reducing the computational cost. Comprehensive experiments verify that our method achieves state-of-the-art performance on single-level and hierarchical text detection.

Significant advancements in video question answering (VideoQA) have been made thanks to thriving large image-language pretraining frameworks. Although these image-language models can efficiently represent both video and language branches they typically employ a goal-free vision perception process and do not interact vision with language well during the answer generation thus omitting crucial visual cues. In this paper we are inspired by the human recognition and learning pattern and propose VideoDistill a framework with language-aware (i.e. goal-driven) behavior in both vision perception and answer generation process. VideoDistill generates answers only from question-related visual embeddings and follows a thinking-observing-answering approach that closely resembles human behavior distinguishing it from previous research. Specifically we develop a language-aware gating mechanism to replace the standard cross-attention avoiding language's direct fusion into visual representations. We incorporate this mechanism into two key components of the entire framework. The first component is a differentiable sparse sampling module which selects frames containing the necessary dynamics and semantics relevant to the questions. The second component is a vision refinement module that merges existing spatial-temporal attention layers to ensure the extraction of multi-grained visual semantics associated with the questions. We conduct experimental evaluations on various challenging video question-answering benchmarks and VideoDistill achieves state-of-the-art performance in both general and long-form VideoQA datasets. In Addition we verify that VideoDistill can effectively alleviate the utilization of language shortcut solutions in the EgoTaskQA dataset.

Although neural networks excel in video action recognition tasks their "black-box" nature makes it challenging to understand the rationale behind their decisions. Recent approaches used inherently interpretable models to analyze video actions in a manner akin to human reasoning. However it has been observed that these interpretable models tend to underperform when compared to their black-box counterparts. In this work we present a new framework called Language-guided Interpretable Action Recognition framework (LaIAR). This framework leverages knowledge from language models to enhance both the recognition capabilities and the interpretability of video models. In essence we reframe the challenge of understanding video model decisions as a task of aligning video and language models. Using the logical reasoning captured by the language model we steer the training of the video model. This integrated approach not only improves the video model's adaptability to different domains but also boosts its overall performance. Extensive experiments on Charades and CAD-120 datasets demonstrate the superior performance and interpretability of our proposed method. The code of LaIAR is available at https://github.com/NingWang2049/LaIAR.

Open-vocabulary querying in 3D space is challenging but essential for scene understanding tasks such as object localization and segmentation. Language-embedded scene representations have made progress by incorporating language features into 3D spaces. However their efficacy heavily depends on neural networks that are resource-intensive in training and rendering. Although recent 3D Gaussians offer efficient and high-quality novel view synthesis directly embedding language features in them leads to prohibitive memory usage and decreased performance. In this work we introduce Language Embedded 3D Gaussians a novel scene representation for open-vocabulary query tasks. Instead of embedding high-dimensional raw semantic features on 3D Gaussians we propose a dedicated quantization scheme that drastically alleviates the memory requirement and a novel embedding procedure that achieves smoother yet high accuracy query countering the multi-view feature inconsistencies and the high-frequency inductive bias in point-based representations. Our comprehensive experiments show that our representation achieves the best visual quality and language querying accuracy across current language-embedded representations while maintaining real-time rendering frame rates on a single desktop GPU.

Recent progress in Large Multimodal Models (LMM) has opened up great possibilities for various applications in the field of human-machine interactions. However developing LMMs that can comprehend reason and plan in complex and diverse 3D environments remains a challenging topic especially considering the demand for understanding permutation-invariant point cloud representations of the 3D scene. Existing works seek help from multi-view images by projecting 2D features to 3D space which inevitably leads to huge computational overhead and performance degradation. In this paper we present LL3DA a Large Language 3D Assistant that takes point cloud as the direct input and responds to both text instructions and visual interactions. The additional visual interaction enables LMMs to better comprehend human interactions with the 3D environment and further remove the ambiguities within plain texts. Experiments show that LL3DA achieves remarkable results and surpasses various 3D vision-language models on both 3D Dense Captioning and 3D Question Answering.

Long video question answering is a challenging task that involves recognizing short-term activities and reasoning about their fine-grained relationships. State-of-the-art video Large Language Models (vLLMs) hold promise as a viable solution due to their demonstrated emergent capabilities on new tasks. However despite being trained on millions of short seconds-long videos vLLMs are unable to understand minutes-long videos and accurately answer questions about them. To address this limitation we propose a lightweight and self-supervised approach Key frame-conditioned long video-LLM (Koala) that introduces learnable spatiotemporal queries to adapt pretrained vLLMs for generalizing to longer videos. Our approach introduces two new tokenizers that condition on visual tokens computed from sparse video key frames for understanding short and long video moments. We train our proposed approach on HowTo100M and demonstrate its effectiveness on zero-shot long video understanding benchmarks where it outperforms state-of-the-art large models by 3 - 6% in absolute accuracy across all tasks. Surprisingly we also empirically show that our approach not only helps a pretrained vLLM to understand long videos but also improves its accuracy on short-term action recognition.

Short-form UGC video platforms like Kwai and TikTok have been an emerging and irreplaceable mainstream media form thriving on user-friendly engagement and kaleidoscope creation etc. However the advancing content generation modes e.g. special effects and sophisticated processing workflows e.g. de-artifacts have introduced significant challenges to recent UGC video quality assessment: (i) the ambiguous contents hinder the identification of quality-determined regions. (ii) the diverse and complicated hybrid distortions are hard to distinguish. To tackle the above challenges and assist in the development of short-form videos we establish the first large-scale Kwai short Video database for Quality assessment termed KVQ which comprises 600 user-uploaded short videos and 3600 processed videos through the diverse practical processing workflows including pre-processing transcoding and enhancement. Among them the absolute quality score of each video and partial ranking score among indistinguish samples are provided by a team of professional researchers specializing in image processing. Based on this database we propose the first short-form video quality evaluator i.e. KSVQE which enables the quality evaluator to identify the quality-determined semantics with the content understanding of large vision language models (i.e. CLIP) and distinguish the distortions with the distortion under- standing module. Experimental results have shown the effectiveness of KSVQE on our KVQ database and popular VQA databases. The project can be found at https: //lixinustc.github.io/projects/KVQ/.

Video transformers have become the de facto standard for human action recognition yet their exclusive reliance on the RGB modality still limits their adoption in certain domains. One such domain is Activities of Daily Living (ADL) where RGB alone is not sufficient to distinguish between visually similar actions or actions observed from multiple viewpoints. To facilitate the adoption of video transformers for ADL we hypothesize that the augmentation of RGB with human pose information known for its sensitivity to fine-grained motion and multiple viewpoints is essential. Consequently we introduce the first Pose Induced Video Transformer: PI-ViT (or π-ViT) a novel approach that augments the RGB representations learned by video transformers with 2D and 3D pose information. The key elements of π-ViT are two plug-in modules 2D Skeleton Induction Module and 3D Skeleton Induction Module that are responsible for inducing 2D and 3D pose information into the RGB representations. These modules operate by performing pose-aware auxiliary tasks a design choice that allows π-ViT to discard the modules during inference. Notably π-ViT achieves the state-of-the-art performance on three prominent ADL datasets encompassing both real-world and large-scale RGB-D datasets without requiring poses or additional computational overhead at inference.

Diffusion models have emerged as the de facto paradigm for video generation. However their reliance on web-scale data of varied quality often yields results that are visually unappealing and misaligned with the textual prompts. To tackle this problem we propose InstructVideo to instruct text-to-video diffusion models with human feedback by reward fine-tuning. InstructVideo has two key ingredients: 1) To ameliorate the cost of reward fine-tuning induced by generating through the full DDIM sampling chain we recast reward fine-tuning as editing. By leveraging the diffusion process to corrupt a sampled video InstructVideo requires only partial inference of the DDIM sampling chain reducing fine-tuning cost while improving fine-tuning efficiency. 2) To mitigate the absence of a dedicated video reward model for human preferences we repurpose established image reward models e.g. HPSv2. To this end we propose Segmental Video Reward a mechanism to provide reward signals based on segmental sparse sampling and Temporally Attenuated Reward a method that mitigates temporal modeling degradation during fine-tuning. Extensive experiments both qualitative and quantitative validate the practicality and efficacy of using image reward models in InstructVideo significantly enhancing the visual quality of generated videos without compromising generalization capabilities. Code and models can be accessed through our project page https://instructvideo.github.io/.