CVPR 2024 Papers

Multimodal large language models (MLLMs) have recently achieved impressive general-purpose vision-language capabilities through visual instruction tuning. However current MLLMs primarily focus on image-level or box-level understanding falling short in achieving fine-grained vision-language alignment at pixel level. Besides the lack of mask-based instruction data limits their advancements. In this paper we propose Osprey a mask-text instruction tuning approach to extend MLLMs by incorporating fine-grained mask regions into language instruction aiming at achieving pixel-wise visual understanding. To achieve this goal we first meticulously curate a mask-based region-text dataset with 724K samples and then design a vision-language model by injecting pixel-level representation into LLM. Specifically Osprey adopts a convolutional CLIP backbone as the vision encoder and employs a mask-aware visual extractor to extract precise visual mask features from high resolution input. Experimental results demonstrate Osprey's superiority in various region understanding tasks showcasing its new capability for pixel-level instruction tuning. In particular Osprey can be integrated with Segment Anything Model (SAM) seamlessly to obtain multi-granularity semantics. The source code dataset and demo can be found at https://github.com/CircleRadon/Osprey.

Event-based semantic segmentation (ESS) is a fundamental yet challenging task for event camera sensing. The difficulties in interpreting and annotating event data limit its scalability. While domain adaptation from images to event data can help to mitigate this issue there exist data representational differences that require additional effort to resolve. In this work for the first time we synergize information from image text and event-data domains and introduce OpenESS to enable scalable ESS in an open-world annotation-efficient manner. We achieve this goal by transferring the semantically rich CLIP knowledge from image-text pairs to event streams. To pursue better cross-modality adaptation we propose a frame-to-event contrastive distillation and a text-to-event semantic consistency regularization. Experimental results on popular ESS benchmarks showed our approach outperforms existing methods. Notably we achieve 53.93% and 43.31% mIoU on DDD17 and DSEC-Semantic without using either event or frame labels.

Current video anomaly detection (VAD) approaches with weak supervisions are inherently limited to a closed-set setting and may struggle in open-world applications where there can be anomaly categories in the test data unseen during training. A few recent studies attempt to tackle a more realistic setting open-set VAD which aims to detect unseen anomalies given seen anomalies and normal videos. However such a setting focuses on predicting frame anomaly scores having no ability to recognize the specific categories of anomalies despite the fact that this ability is essential for building more informed video surveillance systems. This paper takes a step further and explores open-vocabulary video anomaly detection (OVVAD) in which we aim to leverage pre-trained large models to detect and categorize seen and unseen anomalies. To this end we propose a model that decouples OVVAD into two mutually complementary tasks - class-agnostic detection and class-specific classification - and jointly optimizes both tasks. Particularly we devise a semantic knowledge injection module to introduce semantic knowledge from large language models for the detection task and design a novel anomaly synthesis module to generate pseudo unseen anomaly videos with the help of large vision generation models for the classification task. These semantic knowledge and synthesis anomalies substantially extend our model's capability in detecting and categorizing a variety of seen and unseen anomalies. Extensive experiments on three widely-used benchmarks demonstrate our model achieves state-of-the-art performance on OVVAD task.

We study the underexplored but fundamental vision problem of machine understanding of abstract freehand scene sketches. We introduce a sketch encoder that results in semantically-aware feature space which we evaluate by testing its performance on a semantic sketch segmentation task. To train our model we rely only on the availability of bitmap sketches with their brief captions and do not require any pixel-level annotations. To obtain generalization to a large set of sketches and categories we build on a vision transformer encoder pretrained with the CLIP model. We freeze the text encoder and perform visual-prompt tuning of the visual encoder branch while introducing a set of critical modifications. Firstly we augment the classical key-query (k-q) self-attention blocks with value-value (v-v) self-attention blocks. Central to our model is a two-level hierarchical network design that enables efficient semantic disentanglement: The first level ensures holistic scene sketch encoding and the second level focuses on individual categories. We then in the second level of the hierarchy introduce a cross-attention between textual and visual branches. Our method outperforms zero-shot CLIP pixel accuracy of segmentation results by 37 points reaching an accuracy of 85.5% on the FS-COCO sketch dataset. Finally we conduct a user study that allows us to identify further improvements needed over our method to reconcile machine and human understanding of scene sketches.

Multimodal large language models (MLLMs) have gained significant attention due to their strong multimodal understanding capability. However existing works rely heavily on modality-specific encoders which usually differ in architecture and are limited to common modalities. In this paper we present OneLLM an MLLM that aligns eight modalities to language using a unified framework. We achieve this through a unified multimodal encoder and a progressive multimodal alignment pipeline. In detail we first train an image projection module to connect a vision encoder with LLM. Then we build a universal projection module (UPM) by mixing multiple image projection modules and dynamic routing. Finally we progressively align more modalities to LLM with the UPM. To fully leverage the potential of OneLLM in following instructions we also curated a comprehensive multimodal instruction dataset including 2M items from image audio video point cloud depth/normal map IMU and fMRI brain activity. OneLLM is evaluated on 25 diverse benchmarks encompassing tasks such as multimodal captioning question answering and reasoning where it delivers excellent performance. Code data model and online demo are available at https://github.com/csuhan/OneLLM

Frechet Video Distance (FVD) a prominent metric for evaluating video generation models is known to conflict with human perception occasionally. In this paper we aim to explore the extent of FVD's bias toward frame quality over temporal realism and identify its sources. We first quantify the FVD's sensitivity to the temporal axis by decoupling the frame and motion quality and find that the FVD only increases slightly with larger temporal corruption. We then analyze the generated videos and show that via careful sampling from a large set of generated videos that do not contain motions one can drastically decrease FVD without improving the temporal quality. Both studies suggest FVD's basis towards the quality of individual frames. We show that FVD with features extracted from the recent large-scale self-supervised video models is less biased toward image quality. Finally we revisit a few real-world examples to validate our hypothesis.

The core of video understanding tasks such as recognition captioning and tracking is to automatically detect objects or actions in a video and analyze their temporal evolution. Despite sharing a common goal different tasks often rely on distinct model architectures and annotation formats. In contrast natural language processing benefits from a unified output space i.e. text sequences which simplifies the training of powerful foundational language models such as GPT-3 with extensive training corpora. Inspired by this we seek to unify the output space of video understanding tasks by using languages as labels and additionally introducing time and box tokens. In this way a variety of video tasks could be formulated as video-grounded token generation. This enables us to address various types of video tasks including classification (such as action recognition) captioning (covering clip captioning video question answering and dense video captioning) and localization tasks (such as visual object tracking) within a fully shared encoder-decoder architecture following a generative framework. Through comprehensive experiments we demonstrate such a simple and straightforward idea is quite effective and can achieve state-of-the-art or competitive results on seven video benchmarks providing a novel perspective for more universal video understanding. Code is available at \href https://github.com/wangjk666/OmniVid https://github.com/wangjk666/OmniVid .

Due to the resource-intensive nature of training vision-language models on expansive video data a majority of studies have centered on adapting pre-trained image-language models to the video domain. Dominant pipelines propose to tackle the visual discrepancies with additional temporal learners while overlooking the substantial discrepancy for web-scaled descriptive narratives and concise action category names leading to less distinct semantic space and potential performance limitations. In this work we prioritize the refinement of text knowledge to facilitate generalizable video recognition. To address the limitations of the less distinct semantic space of category names we prompt a large language model (LLM) to augment action class names into Spatio-Temporal Descriptors thus bridging the textual discrepancy and serving as a knowledge base for general recognition. Moreover to assign the best descriptors with different video instances we propose Optimal Descriptor Solver forming the video recognition problem as solving the optimal matching flow across frame-level representations and descriptors. Comprehensive evaluations in zero-shot few-shot and fully supervised video recognition highlight the effectiveness of our approach. Our best model achieves a state-of-the-art zero-shot accuracy of 75.1% on Kinetics-600.

The emerging conditional coding-based neural video codec (NVC) shows superiority over commonly-used residual coding-based codec and the latest NVC already claims to outperform the best traditional codec. However there still exist critical problems blocking the practicality of NVC. In this paper we propose a powerful conditional coding-based NVC that solves two critical problems via feature modulation. The first is how to support a wide quality range in a single model. Previous NVC with this capability only supports about 3.8 dB PSNR range on average. To tackle this limitation we modulate the latent feature of the current frame via the learnable quantization scaler. During the training we specially design the uniform quantization parameter sampling mechanism to improve the harmonization of encoding and quantization. This results in a better learning of the quantization scaler and helps our NVC support about 11.4 dB PSNR range. The second is how to make NVC still work under a long prediction chain. We expose that the previous SOTA NVC has an obvious quality degradation problem when using a large intra-period setting. To this end we propose modulating the temporal feature with a periodically refreshing mechanism to boost the quality. Notably under single intra-frame setting our codec can achieve 29.7% bitrate saving over previous SOTA NVC with 16% MACs reduction. Our codec serves as a notable landmark in the journey of NVC evolution. The codes are at https://github.com/microsoft/DCVC.

Our understanding of the generalization capabilities of neural networks NNs is still incomplete. Prevailing explanations are based on implicit biases of gradient descent GD but they cannot account for the capabilities of models from gradientfree methods nor the simplicity bias recently observed in untrained networks This paper seeks other sources of generalization in NNs. To understand the inductive biases provided by architectures independently from GD we examine untrained randomweight networks Even simple MLPs show strong inductive biases uniform sampling in weight space yields a very biased distribution of functions in terms of complexity But unlike common wisdom NNs do not have an inherent simplicity bias This property depends on components such as ReLUs residual connections and layer normalizations Alternative architectures can be built with a bias for any level of complexity. Transformers also inherit all these properties from their building blocks. We provide a fresh explanation for the success of deep learning independent from gradientbased training It points at promising avenues for controlling the solutions implemented by trained models.

Video scene detection aims to temporally link shots for obtaining semantically compact scenes. It is essential for this task to capture scene-distinguishable affinity among shots by similarity assessment. However most methods relies on ordinary shot-to-shot similarities which may inveigle similar shots into being linked even though they are from different scenes and meanwhile hinder dissimilar shots from being blended into a complete scene. In this paper we propose NeighborNet to inject shot contexts into shot-to-shot similarities through carefully exploring the relations between semantic/temporal neighbors of shots over a local time period. In this way shot-to-shot similarities are remeasured as semantic/temporal neighbor-aware similarities so that NeighborNet can learn context embedding into shot features using graph convolutional network. As a result not only do the learned shot features suppress the affinity among similar shots from different scenes but they also promote the affinity among dissimilar shots in the same scene. Experimental results on public benchmark datasets show that our proposed NeighborNet yields substantial improvements in video scene detection especially outperforms released state-of-the-arts by at least 6% in Average Precision (AP). The code is available at https://github.com/ExMorgan-Alter/NeighborNet.

We propose to improve transformers of a specific modality with irrelevant data from other modalities e.g. improve an ImageNet model with audio or point cloud datasets. We would like to highlight that the data samples of the target modality are irrelevant to the other modalities which distinguishes our method from other works utilizing paired (e.g. CLIP) or interleaved data of different modalities. We propose a methodology named Multimodal Pathway - given a target modality and a transformer designed for it we use an auxiliary transformer trained with data of another modality and construct pathways to connect components of the two models so that data of the target modality can be processed by both models. In this way we utilize the universal sequence-to-sequence modeling abilities of transformers obtained from two modalities. As a concrete implementation we use a modality-specific tokenizer and task-specific head as usual but utilize the transformer blocks of the auxiliary model via a proposed method named Cross-Modal Re-parameterization which exploits the auxiliary weights without any inference costs. On the image point cloud video and audio recognition tasks we observe significant and consistent performance improvements with irrelevant data from other modalities. The code and models are available at https://github.com/AILab-CVC/M2PT.

We present MultiPly a novel framework to reconstruct multiple people in 3D from monocular in-the-wild videos. Reconstructing multiple individuals moving and interacting naturally from monocular in-the-wild videos poses a challenging task. Addressing it necessitates precise pixel-level disentanglement of individuals without any prior knowledge about the subjects. Moreover it requires recovering intricate and complete 3D human shapes from short video sequences intensifying the level of difficulty. To tackle these challenges we first define a layered neural representation for the entire scene composited by individual human and background models. We learn the layered neural representation from videos via our layer-wise differentiable volume rendering. This learning process is further enhanced by our hybrid instance segmentation approach which combines the self-supervised 3D segmentation and the promptable 2D segmentation module yielding reliable instance segmentation supervision even under close human interaction. A confidence-guided optimization formulation is introduced to optimize the human poses and shape/appearance alternately. We incorporate effective objectives to refine human poses via photometric information and impose physically plausible constraints on human dynamics leading to temporally consistent 3D reconstructions with high fidelity. The evaluation of our method shows the superiority over prior art on publicly available datasets and in-the-wild videos.

Recently integrating video foundation models and large language models to build a video understanding system can overcome the limitations of specific pre-defined vision tasks. Yet existing systems can only handle videos with very few frames. For long videos the computation complexity memory cost and long-term temporal connection impose additional challenges. Taking advantage of the Atkinson-Shiffrin memory model with tokens in Transformers being employed as the carriers of memory in combination with our specially designed memory mechanism we propose the MovieChat to overcome these challenges. MovieChat achieves state-of-the-art performance in long video understanding along with the released MovieChat-1K benchmark with 1K long video and 14K manual annotations for validation of the effectiveness of our method. The code models and data can be found in https://rese1f.github.io/MovieChat.

Existing diffusion-based video editing models have made gorgeous advances for editing attributes of a source video over time but struggle to manipulate the motion information while preserving the original protagonist's appearance and background. To address this we propose MotionEditor the first diffusion model for video motion editing. MotionEditor incorporates a novel content-aware motion adapter into ControlNet to capture temporal motion correspondence. While ControlNet enables direct generation based on skeleton poses it encounters challenges when modifying the source motion in the inverted noise due to contradictory signals between the noise (source) and the condition (reference). Our adapter complements ControlNet by involving source content to transfer adapted control signals seamlessly. Further we build up a two-branch architecture (a reconstruction branch and an editing branch) with a high-fidelity attention injection mechanism facilitating branch interaction. This mechanism enables the editing branch to query the key and value from the reconstruction branch in a decoupled manner making the editing branch retain the original background and protagonist appearance. We also propose a skeleton alignment algorithm to address the discrepancies in pose size and position. Experiments demonstrate the promising motion editing ability of MotionEditor both qualitatively and quantitatively. To the best of our knowledge MotionEditor is the first to use diffusion models specifically for video motion editing considering the origin dynamic background and camera movement.

Neural rendering has demonstrated remarkable success in dynamic scene reconstruction. Thanks to the expressiveness of neural representations prior works can accurately capture the motion and achieve high-fidelity reconstruction of the target object. Despite this real-world video scenarios often feature large unobserved regions where neural representations struggle to achieve realistic completion. To tackle this challenge we introduce MorpheuS a framework for dynamic 360deg surface reconstruction from a casually captured RGB-D video. Our approach models the target scene as a canonical field that encodes its geometry and appearance in conjunction with a deformation field that warps points from the current frame to the canonical space. We leverage a view-dependent diffusion prior and distill knowledge from it to achieve realistic completion of unobserved regions. Experimental results on various real-world and synthetic datasets show that our method can achieve high-fidelity 360deg surface reconstruction of a deformable object from a monocular RGB-D video.

Undoubtedly high-fidelity 3D hair is crucial for achieving realism artistic expression and immersion in computer graphics. While existing 3D hair modeling methods have achieved impressive performance the challenge of achieving high-quality hair reconstruction persists: they either require strict capture conditions making practical applications difficult or heavily rely on learned prior data obscuring fine-grained details in images. To address these challenges we propose MonoHair a generic framework to achieve high-fidelity hair reconstruction from a monocular video without specific requirements for environments. Our approach bifurcates the hair modeling process into two main stages: precise exterior reconstruction and interior structure inference. The exterior is meticulously crafted using our Patch-based Multi-View Optimization PMVO. This method strategically collects and integrates hair information from multiple views independent of prior data to produce a high-fidelity exterior 3D line map. This map not only captures intricate details but also facilitates the inference of the hair's inner structure. For the interior we employ a data-driven multi-view 3D hair reconstruction method. This method utilizes 2D structural renderings derived from the reconstructed exterior mirroring the synthetic 2D inputs used during training. This alignment effectively bridges the domain gap between our training data and real-world data thereby enhancing the accuracy and reliability of our interior structure inference. Lastly we generate a strand model and resolve the directional ambiguity by our hair growth algorithm. Our experiments demonstrate that our method exhibits robustness across diverse hairstyles and achieves state-of-the-art performance. For more results please refer to our project page https://keyuwu-cs.github.io/MonoHair/

Blind video quality assessment (BVQA) plays a pivotal role in evaluating and improving the viewing experience of end-users across a wide range of video-based platforms and services. Contemporary deep learning-based models primarily analyze video content in its aggressively subsampled format while being blind to the impact of the actual spatial resolution and frame rate on video quality. In this paper we propose a modular BVQA model and a method of training it to improve its modularity. Our model comprises a base quality predictor a spatial rectifier and a temporal rectifier responding to the visual content and distortion spatial resolution and frame rate changes on video quality respectively. During training spatial and temporal rectifiers are dropped out with some probabilities to render the base quality predictor a standalone BVQA model which should work better with the rectifiers. Extensive experiments on both professionally-generated content and user-generated content video databases show that our quality model achieves superior or comparable performance to current methods. Additionally the modularity of our model offers an opportunity to analyze existing video quality databases in terms of their spatial and temporal complexity.

This paper addresses the task of video question answering (videoQA) via a decomposed multi-stage modular reasoning framework. Previous modular methods have shown promise with a single planning stage ungrounded in visual content. However through a simple and effective baseline we find that such systems can lead to brittle behavior in practice for challenging videoQA settings. Thus unlike traditional single-stage planning methods we propose a multi-stage system consisting of an event parser a grounding stage and a final reasoning stage in conjunction with an external memory. All stages are training-free and performed using few-shot prompting of large models creating interpretable intermediate outputs at each stage. By decomposing the underlying planning and task complexity our method MoReVQA improves over prior work on standard videoQA benchmarks (NExT-QA iVQA EgoSchema and ActivityNet-QA) with state-of-the-art results and extensions to related tasks (grounded videoQA paragraph captioning).