NeurIPS 2024 Papers

Diffusion-based Video Super-Resolution (VSR) is renowned for generating perceptually realistic videos, yet it grapples with maintaining detail consistency across frames due to stochastic fluctuations. The traditional approach of pixel-level alignment is ineffective for diffusion-processed frames because of iterative disruptions. To overcome this, we introduce SeeClear--a novel VSR framework leveraging conditional video generation, orchestrated by instance-centric and channel-wise semantic controls. This framework integrates a Semantic Distiller and a Pixel Condenser, which synergize to extract and upscale semantic details from low-resolution frames. The Instance-Centric Alignment Module (InCAM) utilizes video-clip-wise tokens to dynamically relate pixels within and across frames, enhancing coherency. Additionally, the Channel-wise Texture Aggregation Memory (CaTeGory) infuses extrinsic knowledge, capitalizing on long-standing semantic textures. Our method also innovates the blurring diffusion process with the ResShift mechanism, finely balancing between sharpness and diffusion effects. Comprehensive experiments confirm our framework's advantage over state-of-the-art diffusion-based VSR techniques.

Understanding human emotions is fundamental to enhancing human-computer interaction, especially for embodied agents that mimic human behavior. Traditional emotion analysis often takes a third-person perspective, limiting the ability of agents to interact naturally and empathetically. To address this gap, this paper presents $E^3$ for Exploring Embodied Emotion, the first massive first-person view video dataset. $E^3$ contains more than $50$ hours of video, capturing $8$ different emotion types in diverse scenarios and languages. The dataset features videos recorded by individuals in their daily lives, capturing a wide range of real-world emotions conveyed through visual, acoustic, and textual modalities. By leveraging this dataset, we define $4$ core benchmark tasks - emotion recognition, emotion classification, emotion localization, and emotion reasoning - supported by more than $80$k manually crafted annotations, providing a comprehensive resource for training and evaluating emotion analysis models. We further present Emotion-LlaMa, which complements visual modality with acoustic modality to enhance the understanding of emotion in first-person videos. The results of comparison experiments with a large number of baselines demonstrate the superiority of Emotion-LlaMa and set a new benchmark for embodied emotion analysis. We expect that $E^3$ can promote advances in multimodal understanding, robotics, and augmented reality, and provide a solid foundation for the development of more empathetic and context-aware embodied agents.

Despite significant advancements in video generation and editing using diffusion models, achieving accurate and localized video editing remains a substantial challenge. Additionally, most existing video editing methods primarily focus on altering visual content, with limited research dedicated to motion editing. In this paper, we present a novel attempt to Remake a Video (ReVideo) which stands out from existing methods by allowing precise video editing in specific areas through the specification of both content and motion. Content editing is facilitated by modifying the first frame, while the trajectory-based motion control offers an intuitive user interaction experience. ReVideo addresses a new task involving the coupling and training imbalance between content and motion control. To tackle this, we develop a three-stage training strategy that progressively decouples these two aspects from coarse to fine. Furthermore, we propose a spatiotemporal adaptive fusion module to integrate content and motion control across various sampling steps and spatial locations. Extensive experiments demonstrate that our ReVideo has promising performance on several accurate video editing applications, i.e., (1) locally changing video content while keeping the motion constant, (2) keeping content unchanged and customizing new motion trajectories, (3) modifying both content and motion trajectories. Our method can also seamlessly extend these applications to multi-area editing without specific training, demonstrating its flexibility and robustness.

As the scale of data and models for video understanding rapidly expand, handling long-form video input in transformer-based models presents a practical challenge. Rather than resorting to input sampling or token dropping, which may result in information loss, token merging shows promising results when used in collaboration with transformers. However, the application of token merging for long-form video processing is not trivial. We begin with the premise that token merging should not rely solely on the similarity of video tokens; the saliency of tokens should also be considered. To address this, we explore various video token merging strategies for long-form video classification, starting with a simple extension of image token merging, moving to region-concentrated merging, and finally proposing a learnable video token merging (VTM) algorithm that dynamically merges tokens based on their saliency. Extensive experimental results show that we achieve better or comparable performances on the LVU, COIN, and Breakfast datasets. Moreover, our approach significantly reduces memory costs by 84% and boosts throughput by approximately 6.89 times compared to baseline algorithms.

Transferring visual-language knowledge from large-scale foundation models for video recognition has proved to be effective. To bridge the domain gap, additional parametric modules are added to capture the temporal information. However, zero-shot generalization diminishes with the increase in the number of specialized parameters, making existing works a trade-off between zero-shot and close-set performance. In this paper, we present MoTE, a novel framework that enables generalization and specialization to be balanced in one unified model. Our approach tunes a mixture of temporal experts to learn multiple task views with various degrees of data fitting. To maximally preserve the knowledge of each expert, we propose Weight Merging Regularization, which regularizes the merging process of experts in weight space. Additionally with temporal feature modulation to regularize the contribution of temporal feature during test. We achieve a sound balance between zero-shot and close-set video recognition tasks and obtain state-of-the-art or competitive results on various datasets, including Kinetics-400 \& 600, UCF, and HMDB. Code is available at https://github.com/ZMHH-H/MoTE.

Video Anomaly Detection (VAD) systems can autonomously monitor and identify disturbances, reducing the need for manual labor and associated costs. However, current VAD systems are often limited by their superficial semantic understanding of scenes and minimal user interaction. Additionally, the prevalent data scarcity in existing datasets restricts their applicability in open-world scenarios. In this paper, we introduce HAWK, a novel framework that leverages interactive large Visual Language Models (VLM) to interpret video anomalies precisely. Recognizing the difference in motion information between abnormal and normal videos, HAWK explicitly integrates motion modality to enhance anomaly identification. To reinforce motion attention, we construct an auxiliary consistency loss within the motion and video space, guiding the video branch to focus on the motion modality. Moreover, to improve the interpretation of motion-to-language, we establish a clear supervisory relationship between motion and its linguistic representation. Furthermore, we have annotated over 8,000 anomaly videos with language descriptions, enabling effective training across diverse open-world scenarios, and also created 8,000 question-answering pairs for users' open-world questions. The final results demonstrate that HAWK achieves SOTA performance, surpassing existing baselines in both video description generation and question-answering. Our codes/dataset/demo will be released at https://github.com/jqtangust/hawk.

A mainstream of Multi-modal Large Language Models (MLLMs) have two essential functions, i.e., visual recognition (e.g., grounding) and understanding (e.g., visual question answering). Presently, all these MLLMs integrate visual recognition and understanding in a same sequential manner in the LLM head, i.e., generating the response token-by-token for both recognition and understanding. We think unifying them in the same sequential manner is not optimal for two reasons: 1) parallel recognition is more efficient than sequential recognition and is actually prevailing in deep visual recognition, and 2) the recognition results can be integrated to help high-level cognition (while the current manner does not). Such motivated, this paper proposes a novel “parallel recognition → sequential understanding” framework for MLLMs. The bottom LLM layers are utilized for parallel recognition and the recognition results are relayed into the top LLM layers for sequential understanding. Specifically, parallel recognition in the bottom LLM layers is implemented via object queries, a popular mechanism in DEtection TRansformer, which we find to harmonize well with the LLM layers. Empirical studies show our MLLM named Octopus improves accuracy on popular MLLM tasks and is up to 5× faster on visual grounding tasks.

Recent DETR-based methods have advanced the development of Video Instance Segmentation (VIS) through transformers' efficiency and capability in modeling spatial and temporal information. Despite harvesting remarkable progress, existing works follow asynchronous designs, which model video sequences via either video-level queries only or adopting query-sensitive cascade structures, resulting in difficulties when handling complex and challenging video scenarios. In this work, we analyze the cause of this phenomenon and the limitations of the current solutions, and propose to conduct synchronized modeling via a new framework named SyncVIS. Specifically, SyncVIS explicitly introduces video-level query embeddings and designs two key modules to synchronize video-level query with frame-level query embeddings: a synchronized video-frame modeling paradigm and a synchronized embedding optimization strategy. The former attempts to promote the mutual learning of frame- and video-level embeddings with each other and the latter divides large video sequences into small clips for easier optimization. Extensive experimental evaluations are conducted on the challenging YouTube-VIS 2019 & 2021 & 2022, and OVIS benchmarks, and SyncVIS achieves state-of-the-art results, which demonstrates the effectiveness and generality of the proposed approach. The code is available at https://github.com/rkzheng99/SyncVIS.

Speech serves as a ubiquitous input interface for embedded mobile devices. Cloud-based solutions, while offering powerful speech understanding services, raise significant concerns regarding user privacy. To address this, disentanglement-based encoders have been proposed to remove sensitive information from speech signals without compromising the speech understanding functionality. However, these encoders demand high memory usage and computation complexity, making them impractical for resource-constrained wimpy devices. Our solution is based on a key observation that speech understanding hinges on long-term dependency knowledge of the entire utterance, in contrast to privacy-sensitive elements that are short-term dependent. Exploiting this observation, we propose SILENCE, a lightweight system that selectively obscuring short-term details, without damaging the long-term dependent speech understanding performance. The crucial part of SILENCE is a differential mask generator derived from interpretable learning to automatically configure the masking process. We have implemented SILENCE on the STM32H7 microcontroller and evaluate its efficacy under different attacking scenarios. Our results demonstrate that SILENCE offers speech understanding performance and privacy protection capacity comparable to existing encoders, while achieving up to 53.3$\times$ speedup and 134.1$\times$ reduction in memory footprint.

Recent advancements in 3D Large Language Models (LLMs) have demonstrated promising capabilities for 3D scene understanding. However, previous methods exhibit deficiencies in general referencing and grounding capabilities for intricate scene comprehension. In this paper, we introduce the use of object identifiers and object-centric representations to interact with scenes at the object level. Specifically, we decompose the input 3D scene into a set of object proposals, each assigned a unique identifier token, which enables efficient object referencing and grounding during user-assistant interactions. Given the scarcity of scene-language data, we model the scene embeddings as a sequence of explicit object-level embeddings, derived from semantic-rich 2D or 3D representations. By employing object identifiers, we transform diverse 3D scene-language tasks into a unified question-answering format, facilitating joint training without the need for additional task-specific heads. With minimal fine-tuning on all downstream tasks, our model significantly outperforms existing methods on benchmarks including ScanRefer, Multi3DRefer, Scan2Cap, ScanQA, and SQA3D.

Current vision models typically maintain a fixed correspondence between their representation structure and image space. Each layer comprises a set of tokens arranged “on-the-grid,” which biases patches or tokens to encode information at a specific spatio(-temporal) location. In this work we present *Moving Off-the-Grid* (MooG), a self-supervised video representation model that offers an alternative approach, allowing tokens to move “off-the-grid” to better enable them to represent scene elements consistently, even as they move across the image plane through time. By using a combination of cross-attention and positional embeddings we disentangle the representation structure and image structure. We find that a simple self-supervised objective—next frame prediction—trained on video data, results in a set of latent tokens which bind to specific scene structures and track them as they move. We demonstrate the usefulness of MooG’s learned representation both qualitatively and quantitatively by training readouts on top of the learned representation on a variety of downstream tasks. We show that MooG can provide a strong foundation for different vision tasks when compared to “on-the-grid” baselines.

Adverse weather conditions can significantly degrade the video frames, causing existing video semantic segmentation methods to produce erroneous predictions. In this work, we target adverse weather conditions and introduce an end-to-end domain adaptation strategy that leverages a fusion block, temporal-spatial teacher-student learning, and a temporal weather degradation augmentation approach. The fusion block integrates temporal information from adjacent frames at the feature level, trained end-to-end, eliminating the need for pretrained optical flow, distinguishing our method from existing approaches. Our teacher-student approach involves two teachers: one focuses on exploring temporal information from adjacent frames, and the other harnesses spatial information from the current frame. Finally, we apply temporal weather degradation augmentation to consecutive frames to more accurately represent adverse weather degradations. Our method achieves a performance of 25.4 and 33.0 mIoU on the adaptation from VIPER and Synthia to MVSS, respectively, representing an improvement of 4.3 and 5.8 mIoU over the existing state-of-the-art method.

The arrival of Sora marks a new era for text-to-video diffusion models, bringing significant advancements in video generation and potential applications. However, Sora, along with other text-to-video diffusion models, is highly reliant on prompts, and there is no publicly available dataset that features a study of text-to-video prompts. In this paper, we introduce VidProM, the first large-scale dataset comprising 1.67 Million unique text-to-Video Prompts from real users. Additionally, this dataset includes 6.69 million videos generated by four state-of-the-art diffusion models, alongside some related data. We initially discuss the curation of this large-scale dataset, a process that is both time-consuming and costly. Subsequently, we underscore the need for a new prompt dataset specifically designed for text-to-video generation by illustrating how VidProM differs from DiffusionDB, a large-scale prompt-gallery dataset for image generation. Our extensive and diverse dataset also opens up many exciting new research areas. For instance, we suggest exploring text-to-video prompt engineering, efficient video generation, and video copy detection for diffusion models to develop better, more efficient, and safer models. The project (including the collected dataset VidProM and related code) is publicly available at https://vidprom.github.io under the CC-BY-NC 4.0 License.

Deep Equilibrium Model (DEQ), which serves as a typical implicit neural network, emphasizes their memory efficiency and competitive performance compared to explicit neural networks. However, there has been relatively limited theoretical analysis on the representation of DEQ. In this paper, we utilize the Neural Collapse ($\mathcal{NC}$) as a tool to systematically analyze the representation of DEQ under both balanced and imbalanced conditions. $\mathcal{NC}$ is an interesting phenomenon in the neural network training process that characterizes the geometry of class features and classifier weights. While extensively studied in traditional explicit neural networks, the $\mathcal{NC}$ phenomenon has not received substantial attention in the context of implicit neural networks. We theoretically show that $\mathcal{NC}$ exists in DEQ under balanced conditions. Moreover, in imbalanced settings, despite the presence of minority collapse, DEQ demonstrated advantages over explicit neural networks. These advantages include the convergence of extracted features to the vertices of a simplex equiangular tight frame and self-duality properties under mild conditions, highlighting DEQ's superiority in handling imbalanced datasets. Finally, we validate our theoretical analyses through experiments in both balanced and imbalanced scenarios.

Graph Neural Networks (GNNs) excel across various applications but remain vulnerable to adversarial attacks, particularly Graph Injection Attacks (GIAs), which inject malicious nodes into the original graph and pose realistic threats. Text-attributed graphs (TAGs), where nodes are associated with textual features, are crucial due to their prevalence in real-world applications and are commonly used to evaluate these vulnerabilities. However, existing research only focuses on embedding-level GIAs, which inject node embeddings rather than actual textual content, limiting their applicability and simplifying detection. In this paper, we pioneer the exploration of GIAs at the text level, presenting three novel attack designs that inject textual content into the graph. Through theoretical and empirical analysis, we demonstrate that text interpretability, a factor previously overlooked at the embedding level, plays a crucial role in attack strength. Among the designs we investigate, the Word-frequency-based Text-level GIA (WTGIA) is particularly notable for its balance between performance and interpretability. Despite the success of WTGIA, we discover that defenders can easily enhance their defenses with customized text embedding methods or large language model (LLM)--based predictors. These insights underscore the necessity for further research into the potential and practical significance of text-level GIAs.

Diffusion models have obtained substantial progress in image-to-video generation. However, in this paper, we find that these models tend to generate videos with less motion than expected. We attribute this to the issue called conditional image leakage, where the image-to-video diffusion models (I2V-DMs) tend to over-rely on the conditional image at large time steps. We further address this challenge from both inference and training aspects. First, we propose to start the generation process from an earlier time step to avoid the unreliable large-time steps of I2V-DMs, as well as an initial noise distribution with optimal analytic expressions (Analytic-Init) by minimizing the KL divergence between it and the actual marginal distribution to bridge the training-inference gap. Second, we design a time-dependent noise distribution (TimeNoise) for the conditional image during training, applying higher noise levels at larger time steps to disrupt it and reduce the model's dependency on it. We validate these general strategies on various I2V-DMs on our collected open-domain image benchmark and the UCF101 dataset. Extensive results show that our methods outperform baselines by producing higher motion scores with lower errors while maintaining image alignment and temporal consistency, thereby yielding superior overall performance and enabling more accurate motion control. The project page: \url{https://cond-image-leak.github.io/}.

Outlier Features (OFs) are neurons whose activation magnitudes significantly exceed the average over a neural network's (NN) width. They are well known to emerge during standard transformer training and have the undesirable effect of hindering quantisation in afflicted models. Despite their practical importance, little is known behind why OFs emerge during training, nor how one can minimise them. Our work focuses on the above questions, first identifying several quantitative metrics, such as the kurtosis over neuron activation norms, to measure OFs. With these metrics, we study how architectural and optimisation choices influence OFs, and provide practical insights to minimise OFs during training. As highlights, we introduce a novel unnormalised transformer block, the Outlier Protected block, and present a previously unknown benefit of non-diagonal preconditioning optimisers, finding both approaches to significantly reduce OFs and improve quantisation without compromising convergence speed, at scales of up to 7B parameters. Notably, our combination of OP block and non-diagonal preconditioner (SOAP) achieves 14.87 weight-and-activation int8 perplexity (from 14.71 in standard precision), compared to 63.4 int8 perplexity (from 16.00) with a default OF-prone combination of Pre-Norm model and Adam, when quantising OPT-125m models post-training.

Contrastive Language-Image Pre-training (CLIP) achieves impressive performance on tasks like image classification and image-text retrieval by learning on large-scale image-text datasets. However, CLIP struggles with dense prediction tasks due to the poor grasp of the fine-grained details. Although existing works pay attention to this issue, they achieve limited improvements and usually sacrifice the important visual-semantic consistency. To overcome these limitations, we propose FineCLIP, which keeps the global contrastive learning to preserve the visual-semantic consistency and further enhances the fine-grained understanding through two innovations: 1) A real-time self-distillation scheme that facilitates the transfer of representation capability from global to local features. 2) A semantically-rich regional contrastive learning paradigm with generated region-text pairs, boosting the local representation capabilities with abundant fine-grained knowledge. Both cooperate to fully leverage diverse semantics and multi-grained complementary information. To validate the superiority of our FineCLIP and the rationality of each design, we conduct extensive experiments on challenging dense prediction and image-level tasks. All the observations demonstrate the effectiveness of FineCLIP.

Existing vision-language understanding benchmarks largely consist of images of objects in their usual contexts. As a consequence, recent multimodal large language models can perform well with only a shallow visual understanding by relying on background language biases. Thus, strong performance on these benchmarks does not necessarily correlate with strong visual understanding. In this paper, we release JourneyBench, a comprehensive human-annotated benchmark of generated images designed to assess the model's fine-grained multimodal reasoning abilities across five tasks: complementary multimodal chain of thought, multi-image VQA, imaginary image captioning, VQA with hallucination triggers, and fine-grained retrieval with sample-specific distractors. Unlike existing benchmarks, JourneyBench explicitly requires fine-grained multimodal reasoning in unusual imaginary scenarios where language bias and holistic image gist are insufficient. We benchmark state-of-the-art models on JourneyBench and analyze performance along a number of fine-grained dimensions. Results across all five tasks show that JourneyBench is exceptionally challenging for even the best models, indicating that models' visual reasoning abilities are not as strong as they first appear. We discuss the implications of our findings and propose avenues for further research.