论文检索

Wei Feng, Kangrui Ye, Qi Zhang, Qian Zhang, Nan Li

Gaussian splatting techniques have recently revolutionized outdoor scene decomposition and relighting through multi-view images. However, achieving high rendering quality still requires a fixed lighting condition among all input views, which is costly or even impractical to capture in outdoor scenes. In this paper, we propose outdoor scene decomposition and relighting with 2D Gaussian splatting (OSDR-GS), a novel inverse rendering strategy under outdoor changing and unknown lighting conditions. Firstly, we present a lighting-based group learning framework that categorizes input images into multiple lighting groups, to learn the separate lighting from each group individually. Secondly, OSDR-GS introduces a fine-grained outdoor lighting component to represent sun-light and sky-light, respectively, which are also adjusted via the correlative exposure factors adaptively. Finally, we construct a visibility-driven shadow module to characterize the nuanced interplay of light and occlusion realistically, for eliminating the uncertainty of dark pixels on lighting-based group learning. Extensive experiments on multiple challenging outdoor datasets validate the effectiveness of OSDR-GS, which achieves the state-of-the-art performance in changing lighting scene inverse rendering.

Zhirui Fang, Yi Li, Xin Xie, Chengyan Li, Yanqing Guo

Style transfer is a challenging task in computer vision, aiming to blend the stylistic features of one image with the content of another while preserving the content details. Traditional methods often face challenges in terms of computational efficiency and fine-grained content preservation. In this paper, we propose a novel feature modulation mechanism based on parameterized normalization, where the modulation parameters for content and style features are learned using a dual convolution network (BiConv). These parameters adjust the mean and standard deviation of the features, improving both the stability and quality of the style transfer process. To achieve fast inference, we introduce an efficient acceleration technique by leveraging a row and column weighted attention matrix. In addition, we incorporate a contrastive learning scheme to align the local features of the content and the stylized images, improving the fidelity of the generated output. Experimental results demonstrate that our method significantly improves the inference speed and the quality of style transfer while preserving content details, outperforming existing approaches based on both convolution and diffusion.

Jisheng Dang, Ligen Chen, Jingze Wu, Ronghao Lin, Bimei Wang, Yun Wang, Liting Wang, Nannan Zhu, Teng Wang

Dynamic spatio-temporal understanding is essential for video-based multimodal tasks, yet existing methods often struggle to capture fine-grained temporal and spatial relationships in long videos. Current approaches primarily rely on pre-trained CLIP encoders, which excel in semantic understanding but lack spatially-aware visual context. This leads to hallucinated results when interpreting fine-grained objects or scenes. To address these limitations, we propose a novel framework that integrates diffusion models into multimodal video models. By employing diffusion encoders at intermediate layers, we enhance visual representations through feature alignment and knowledge distillation losses, significantly improving the model's ability to capture spatial patterns over time. Additionally, we introduce a multi-level alignment strategy to learn robust feature correspondence from pre-trained diffusion models. Extensive experiments on benchmark datasets demonstrate our approach's state-of-the-art performance across multiple video understanding tasks. These results establish diffusion models as a powerful tool for enhancing multimodal video models in complex, dynamic scenarios.

Jisheng Dang, Huicheng Zheng, Xudong Wu, Jingmei Jiao, Bimei Wang, Jun Yang, Bin Hu, Jianhuang Lai, Tat Seng Chua

Long video understanding with Large Language Models (LLMs) enables the description of objects that are not explicitly present in the training data. However, continuous changes in known objects and the emergence of new ones require up-to-date knowledge of objects and their dynamics for effective understanding of the open world. To alleviate this, we propose an efficient Retrieval-Enhanced Video Understanding method, dubbed REVU, which leverages external knowledge to enhance the performance of open-world learning. First, REVU introduces an extensible external text-object memory with minimal text-visual mapping, involving static and dynamic multimodal information to help LLMs-based models align text and vision features. Second, REVU retrieves object information from external databases and dynamically integrates frame-specific data from videos, enabling effective knowledge aggregation to comprehend the open world. We conducted experiments on multiple benchmark datasets, and our model demonstrates strong adaptability to out-of-domain data without requiring additional fine-tuning or re-training. Experiments on benchmark video understanding datasets reveal that our model achieves state-of-the-art performance and robust generalization.

Guo Chen, Yifei Huang, Yin-dong Zheng, Yicheng Liu, Jiahao Wang, Tong Lu

Egocentric object-interaction anticipation is critical for applications like augmented reality and robotics, but existing methods struggle with misaligned egocentric encoding, insufficient supervision, and underutilized historical context. These limitations stem from a lack of focus on retention, i.e., retaining long-term object-centric interactions, and prediction, i.e., future-centric encoding and future uncertainty modeling. We introduce EgoAnticipator, a novel Retentive and Predictive Learning framework that addresses these challenges. Our approach combines retentive pre-training for domain-specific encoding, predictive pre-training for future uncertainty modeling, and mirror distillation to transfer future-informed knowledge. Additionally, we propose long-term memory prompting to integrate historical interaction cues. We evaluate the effectiveness of our framework using the Ego4D short-term object interaction anticipation benchmark, covering both STAv1 and STAv2. Extensive experiments demonstrate that our framework outperforms existing methods, while ablation studies highlight the effectiveness of each design inside our retentive and predictive learning framework.

Niaz Ahmad, Jawad Khan, Kang G. Shin, Youngmoon Lee, Guanghui Wang

The dynamic movement of the human body presents a fundamental challenge for human pose estimation and body segmentation. State-of-the-art approaches primarily rely on combining keypoint heatmaps with segmentation masks, but often struggle in scenarios involving overlapping joints during pose estimation or rapidly changing poses for instance-level segmentation. To address these limitations, we leverage Keypoints as Dynamic Centroid (KDC), a new centroid-based representation for unified human pose estimation and instance-level segmentation. KDC adopts a bottom-up paradigm to generate keypoint heatmaps for easily distinguishable and complex keypoints, and improves keypoint detection and confidence scores by introducing KeyCentroids using a keypoint disk. It leverages high-confidence keypoints as dynamic centroids in the embedding space to generate MaskCentroids, allowing for the swift clustering of pixels to specific human instances during rapid changes in human body movements in a live environment. Our experimental evaluations focus on crowded and occluded cases using the CrowdPose, OCHuman, and COCO benchmarks, demonstrating KDC’s effectiveness and generalizability in challenging scenarios in terms of both accuracy and runtime performance. Our implementation is available at https://sites.google.com/view/niazahmad/projects/kdc.

Harsh Dubey, Chulwoo Pack

To address the limitations of current Large-scale Video-Language Models (LVLMs) in fine-grained understanding and long-term temporal memory, we propose a novel video understanding approach that integrates a Vision Language Model (VLM) and a Large Language Model (LLM) with a textual memory mechanism to ensure continuity and contextual coherence. In addition, we introduce a novel evaluation metric, VAD-Score (Video Automated Description Score), to assess precision, recall, and F1 scores for events, subjects, and objects. Our approach delivers competitive results on a diverse set of videos from the DREAM-1K dataset, spanning categories such as live-action, animation, shorts, stock, and YouTube, with a focus on fine-grained comprehension.

David Touretzky, Christina Gardner-McCune, Will Hanna, Angela Chen, Neel Pawar

Neuron Sandbox is a browser-based tool that helps middle school students grasp basic principles of neural computation. It simulates a linear threshold unit applied to binary decision problems, which students solve by adjusting the unit's threshold and/or weights. Although Neuron Sandbox provides extensive visualization aids, solving these problems is challenging for students who have not yet been exposed to algebra. We collected survey, video, and worksheet data from 21 seventh grade students in two sections of an AI elective, taught by the same teacher, that used Neuron Sandbox. We present a scaffolding strategy that proved effective at guiding these students to achieve mastery of these problems. While the amount of scaffolding required was more than we originally anticipated, by the end of the exercise students understood the computation that linear threshold units perform and were able to generalize their understanding of the worksheet’s "solve for threshold" strategy to also solve for weights.

Linchao Zhu

Recent advances in vision-language models have shown remarkable potential, yet creating scalable systems that can effectively understand and generate across modalities remains challenging. This talk will present our contributions to advancing scalable vision-language systems, focusing on three key themes: (1) efficient vision-language understanding, including our work on temporal perceiving video-language pre-training and knowledge-enhanced zero-shot retrieval; (2) scalable generation frameworks, encompassing our innovations in zero-shot captioning and co-speech gesture generation; and (3) practical applications and deployments of these technologies. We will discuss how these advances have enabled both better performance and improved efficiency in real-world scenarios, and explore future directions for scalable multimodal systems.

Yi Feng, Chuanyi Li, Vincent Ng

While AI systems are capable of reading texts and seeing images, they typically perceive surface information explicitly conveyed with limited abilities to comprehend hidden messages (e.g., a double-edged remark). We propose the novel task of advertisement understanding: given an advertisement, which can be a text, an image, or a video, the goal is to identify the persuasion strategies used and determine the (possibly hidden) messages conveyed. Efforts on this task could enhance machine comprehension capabilities, and provide users with increased situation awareness w.r.t. the advertised message and thus possibly enable mindful decision making. We believe that this task presents long-term challenges to AI researchers and that successful understanding of ads could bring machine understanding one important step closer to human understanding.

Archit Kambhamettu, Samantha Snyder, Maliheh Fakhar, Samuel Audia, Ross Miller, Jae Kun Shim, Aniket Bera

Understanding internal joint loading is critical for diagnosing gait-related diseases such as knee osteoarthritis; however, current methods of measuring joint risk factors are time-consuming, expensive, and restricted to lab settings. In this paper, we enable the large-scale, cost-effective biomechanical analysis of joint loading via three key contributions: the development and deployment of novel instrumented insoles, the creation of a large multimodal biomechanics dataset (VidSole), and a baseline deep learning pipeline to predict internal joint loading factors. Our novel instrumented insole measures the tri-axial forces and moments across five high-pressure points under the foot. VidSole consists of the forces and moments measured by these insoles along with corresponding RGB video from two viewpoints, 3D body motion capture, and force plate data for over 2,600 trials of 52 diverse participants performing four fundamental activities of daily living (sit-to-stand, stand-to-sit, walking, and running). We feed the insole data and kinematic parameters extractable from video (i.e., pose, knee angle) into a deep learning pipeline consisting of an ensemble Gated Recurrent Unit (GRU) activity classifier followed by activity-specific Long Short Term Memory (LSTM) regression networks to estimate knee adduction moment (KAM), a biomechanical risk factor for knee osteoarthritis. The successful classification of activities at an accuracy of 99.02 percent and KAM estimation with mean absolute error (MAE) less than 0.5 percent*body weight*height, the current threshold for accurately detecting knee osteoarthritis with KAM, illustrates the usefulness of our dataset for future research and clinical settings.

Yueqian Wang, Xiaojun Meng, Yuxuan Wang, Jianxin Liang, Qun Liu, Dongyan Zhao

Multi-modal multi-party conversation (MMC) is a less studied yet important topic of research due to that it well fits real-world scenarios and thus potentially has more widely-used applications. Compared with the traditional multi-modal conversations, MMC requires stronger character-centered understanding abilities as there are many interlocutors appearing in both the visual and textual context. To facilitate the study of this problem, we present Friends-MMC in this paper, an MMC dataset that contains 24,000+ unique utterances paired with video context. To explore the character-centered understanding of the dialogue, we also annotate the speaker of each utterance, the names and bounding bboxes of faces that appear in the video. Based on this Friends-MMC dataset, we further study two fundamental MMC tasks: conversation speaker identification and conversation response prediction, both of which have the multi-party nature with the video or image as visual context. For conversation speaker identification, we demonstrate the inefficiencies of existing methods such as pre-trained models, and propose a simple yet effective baseline method that leverages an optimization solver to utilize the context of two modalities to achieve better performance. For conversation response prediction, we fine-tune generative dialogue models on Friend-MMC, and analyze the benefits of speaker information. The code and dataset will be publicly available, and thus we call for more attention on modelling speaker information when understanding conversations.

Wiradee Imrattanatrai, Masaki Asada, Kimihiro Hasegawa, Zhi-Qi Cheng, Ken Fukuda, Teruko Mitamura

This paper presents VDAct, a dataset for a Video-grounded Dialogue on Event-driven Activities, alongside VDEval, a session-based context evaluation metric specially designed for the task. Unlike existing datasets, VDAct includes longer and more complex video sequences that depict a variety of event-driven activities that require advanced contextual understanding for accurate response generation. The dataset comprises 3,000 dialogues with over 30,000 question-and-answer pairs, derived from 1,000 videos with diverse activity scenarios. VDAct displays a notably challenging characteristic due to its broad spectrum of activity scenarios and wide range of question types. Empirical studies on state-of-the-art vision foundation models highlight their limitations in addressing certain question types on our dataset. Furthermore, VDEval, which integrates dialogue session history and video content summaries extracted from our supplementary Knowledge Graphs to evaluate individual responses, demonstrates a significantly higher correlation with human assessments on the VDAct dataset than existing evaluation metrics that rely solely on the context of single dialogue turns.

Beibei Zhang, Tongwei Ren, Gangshan Wu

Video speaking style recognition (VSSR) aims to classify different types of conversations in videos, contributing significantly to understanding human interactions. A significant challenge in VSSR is the inherent similarity among conversation videos, which makes it difficult to distinguish between different speaking styles. Existing VSSR methods commit to providing available multimodal information to enhance the differentiation of conversation videos. Nevertheless, treating each modality equally leads to a suboptimal result for these methods due to text is inherently more aligned with conversation understanding compared to nonverbal modalities. To address this issue, we propose a text-guided nonverbal enhancement method, TNvE, which is composed of two core modules: 1) a text-guided nonverbal representation selection module employs cross-modal attention based on modality-invariant representations, picking out critical nonverbal information via textual guide; and 2) a modality-invariant and -specific representation decoupling module incorporates modality-specific representations and decouples them from modality-invariant representations, enabling a more comprehensive understanding of multimodal data. The former module encourages multimodal representations close to each other, while the latter module provides unique characteristics of each modality as a supplement. Extensive experiments are conducted on long-form video understanding datasets to demonstrate that TNvE is highly effective for VSSR, achieving a new state-of-the-art.

Shiyu Wang, Yihao Feng, Tian Lan, Ning Yu, Yu Bai, Ran Xu, Huan Wang, Caiming Xiong, Silvio Savarese

Natural language serves as a common and straightforward control signal for humans to interact seamlessly with machines. Recognizing the importance of this interface, the machine learning community is investing considerable effort in generating data that is semantically coherent with textual instructions. While strides have been made in text-to-data generation spanning image editing, audio synthesis, video creation, and beyond, low-resource areas characterized by expensive annotations or complex data structures, such as molecules, motion dynamics, and time series, often lack textual labels. This deficiency impedes supervised learning, thereby constraining the application of advanced generative models for text-to-data tasks. In response to these challenges in the low-resource scenario, we propose Text2Data, a novel approach that utilizes unlabeled data to understand the underlying data distribution through an unsupervised diffusion model. Subsequently, it undergoes controllable finetuning via a novel constraint optimization-based learning objective that ensures controllability and effectively counteracts catastrophic forgetting. Comprehensive experiments demonstrate that Text2Data is able to achieve enhanced performance regarding controllability across various modalities, including molecules, motions and time series, when compared to existing baselines.

Jean Park, Kuk Jin Jang, Basam Alasaly, Sriharsha Mopidevi, Andrew Zolensky, Eric Eaton, Insup Lee, Kevin Johnson

Multimodal large language models (MLLMs) can simultaneously process visual, textual, and auditory data, capturing insights that complement human analysis. However, existing video question-answering (VidQA) benchmarks and datasets often exhibit a bias toward a single modality, despite the goal of requiring advanced reasoning skills that integrate diverse modalities to answer the queries. In this work, we introduce the modality importance score (MIS) to identify such bias. It is designed to assess which modality embeds the necessary information to answer the question. Additionally, we propose an innovative method using state-of-the-art MLLMs to estimate the modality importance, which can serve as a proxy for human judgments of modality perception. With this MIS, we demonstrate the presence of unimodal bias and the scarcity of genuinely multimodal questions in existing datasets. We further validate the modality importance score with multiple ablation studies to evaluate the performance of MLLMs on permuted feature sets. Our results indicate that current models do not effectively integrate information due to modality imbalance in existing datasets. Our proposed MLLM-derived MIS can guide the curation of modality-balanced datasets that advance multimodal learning and enhance MLLMs' capabilities to understand and utilize synergistic relations across modalities.

Hao Ma, Shijie Wang, Zhiqiang Pu, Siyao Zhao, Xiaolin Ai

Guiding the policy of multi-agent reinforcement learning to align with human common sense is a difficult problem, largely due to the complexity of modeling common sense as a reward, especially in complex and long-horizon multi-agent tasks. Recent works have shown the effectiveness of reward shaping, such as potential-based rewards, to enhance policy alignment. The existing works, however, primarily rely on experts to design rule-based rewards, which are often labor-intensive and lack a high-level semantic understanding of common sense. To solve this problem, we propose a hierarchical vision-based reward shaping method. At the bottom layer, a visual-language model (VLM) serves as a generic potential function, guiding the policy to align with human common sense through its intrinsic semantic understanding. To help the policy adapts to uncertainty and changes in long-horizon tasks, the top layer features an adaptive skill selection module based on a visual large language model (vLLM). The module uses instructions, video replays, and training records to dynamically select suitable potential function from a pre-designed pool. Besides, our method is theoretically proven to preserve the optimal policy. Extensive experiments conducted in the Google Research Football environment demonstrate that our method not only achieves a higher win rate but also effectively aligns the policy with human common sense.

Daoming Zong, Chaoyue Ding, Kaitao Chen, Yinsheng Li, Shuaiyu Wang

Physical commonsense is an essential aspect of human cognition, involving an intuitive understanding of the physical properties and interactions of everyday objects and materials. Though physical commonsense reasoning should inherently be a multisensory task, integrating both video and audio signals, existing physical audiovisual commonsense reasoning (PACR) models predominantly rely on visual information. This reliance leads to spurious correlations and undermines the models’ reasoning and generalization abilities. To counteract this, we introduce a model-agnostic Counterfactual Physical Audiovisual Commonsense Reasoning (CF-PACR) framework aimed at mitigating visual bias-induced spurious effects. Specifically, we construct a traditional PACR model using both audio and visual information as the factual reasoning model. Subsequently, in the counterfactual reasoning model, we isolate visual information to estimate direct effects. Finally, we subtract the direct effects from the total effects across modalities to derive indirect effects, thereby mitigating visual biases. Extensive experiments validate the effectiveness and generalizability of CF-PACR in alleviating the spurious correlations between visual modality and model predictions.

Xun Liang, Zhiying Li, Hongxun Jiang

Cross-domain recommendations in healthcare services differ from traditional ones in electronic commerce due to the need for heightened medical privacy protection for a small group of users, while ensuring the majority, who may lack sufficient medical knowledge, can understand the recommendations. To recommend doctors who provide online consultations to health video viewers and enable multimodal cross-domain recommendations from short video platforms (source domain) to online healthcare communities (target domain), this paper introduces a framework based on the User-Centric Synthetic Data Architect (UCSDA) and Pre-trained Large Language Model (PtLLM). UCSDA employs a user-centric, advanced selection-synthesis mechanism to filter users' cold interaction items and synthesize noise items, reducing privacy leakage risk. PtLLM focuses on necessary patient and doctor IDs during the recommendation decision process to generate explanations. The model's effectiveness and scalability were validated using three public datasets and a healthcare cross-domain recommendation dataset. In addition to traditional evaluation metrics, strong privacy metrics and the unique sentence ratio were used to assess privacy protection and interpretability. We also compared the characteristics of privacy protection and interpretability between e-commerce and healthcare recommendation scenarios.

Bochao Zou, Zizheng Guo, Xiaocheng Hu, Huimin Ma

Remote photoplethysmography (rPPG) is a method for non-contact measurement of physiological signals from facial videos, holding great potential in various applications such as healthcare, affective computing, and anti-spoofing. Existing deep learning methods struggle to address two core issues of rPPG simultaneously: understanding the periodic pattern of rPPG among long contexts and addressing large spatiotemporal redundancy in video segments. These represent a trade-off between computational complexity and the ability to capture long-range dependencies. In this paper, we introduce RhythmMamba, a state space model-based method that captures long-range dependencies while maintaining linear complexity. By viewing rPPG as a time series task through the proposed frame stem, the periodic variations in pulse waves are modeled as state transitions. Additionally, we design multi-temporal constraint and frequency domain feed-forward, both aligned with the characteristics of rPPG time series, to improve the learning capacity of Mamba for rPPG signals. Extensive experiments show that RhythmMamba achieves state-of-the-art performance with 319% throughput and 23% peak GPU memory.