AdaSpark: Adaptive Sparsity for Efficient Long-Video Understanding
School of Artificial Intelligence, University of Chinese Academy of Sciences（中国科学院大学人工智能学院）；Institute of Automation, Chinese Academy of Sciences（中国科学院自动化研究所）；Alibaba Group Holding Limited（阿里巴巴集团控股有限公司）；Future Living Lab of Alibaba（阿里巴巴未来生活实验室）
Handong Li, Zikang Liu, Longteng Guo, Tongtian Yue, Yepeng Tang, Xinxin Zhu, Chuanyang Zheng, Ziming Wang, Zhibin Wang, Jun Song, Cheng Yu, Bo Zheng, Jing Liu
备注：
8 pages, CVPR2026 Accept (Highlight)
发表信息：
Proceedings of the IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2026
Processing long-form videos with Video Large Language Models (Video-LLMs) is computationally prohibitive. Current efficiency methods often compromise fine-grained perception through irreversible information disposal or inhibit long-range temporal modeling via rigid, predefined sparse patterns. This paper introduces AdaSpark, an adaptive sparsity framework designed to address these limitations. AdaSpark first partitions video inputs into 3D spatio-temporal cubes. It then employs two co-designed, context-aware components: (1) Adaptive Cube-Selective Attention (AdaS-Attn), which adaptively selects a subset of relevant video cubes to attend for each query token, and (2) Adaptive Token-Selective FFN (AdaS-FFN), which selectively processes only the most salient tokens within each cube. An entropy-based (Top-p) selection mechanism adaptively allocates computational resources based on input complexity. Experiments demonstrate that AdaSpark significantly reduces computational load by up to 57% FLOPs while maintaining comparable performance to dense models and preserving fine-grained, long-range dependencies, as validated on challenging hour-scale video benchmarks.
Activation sparsity improves compute efficiency and resource utilization in sparsity-aware neural network accelerators. As the predominant operation in DNNs is multiply-accumulate (MAC) of activations with weights to compute inner products, skipping operations where (at least) one of the two operands is zero can make inference more efficient in terms of latency and power. Spatial sparsification of activations is a popular topic in DNN literature and several methods have already been established to bias a DNN for it. On the other hand, temporal sparsity is an inherent feature of bio-inspired sp…
Nanyang Technological University（南洋理工大学）；Institute for Infocomm Research（资讯通信研究院）；City University of Hong Kong（香港城市大学）
The articulated and complex nature of human actions makes the task of action recognition difficult. One approach to handle this complexity is dividing it to the kinetics of body parts and analyzing the actions based on these partial descriptors. We propose a joint sparse regression based learning method which utilizes the structured sparsity to model each action as a combination of multimodal features from a sparse set of body parts. To represent dynamics and appearance of parts, we employ a heterogeneous set of depth and skeleton based features. The proper structure of multimodal multipart fe…
We propose Spatio-temporal Crop Aggregation for video representation LEarning (SCALE), a novel method that enjoys high scalability at both training and inference time. Our model builds long-range video features by learning from sets of video clip-level features extracted with a pre-trained backbone. To train the model, we propose a self-supervised objective consisting of masked clip feature prediction. We apply sparsity to both the input, by extracting a random set of video clips, and to the loss function, by only reconstructing the sparse inputs. Moreover, we use dimensionality reduction by w…
Michigan State University（密歇根州立大学）；Georgia Institute of Technology（佐治亚理工学院）；UC, Santa Barbara（加州大学圣巴巴拉分校）；University of Texas at Austin（德克萨斯大学奥斯汀分校）
In this paper, we propose a data-model-hardware tri-design framework for high-throughput, low-cost, and high-accuracy multi-object tracking (MOT) on High-Definition (HD) video stream. First, to enable ultra-light video intelligence, we propose temporal frame-filtering and spatial saliency-focusing approaches to reduce the complexity of massive video data. Second, we exploit structure-aware weight sparsity to design a hardware-friendly model compression method. Third, assisted with data and model complexity reduction, we propose a sparsity-aware, scalable, and low-power accelerator design, aimi…
VideoMem: Enhancing Ultra-Long Video Understanding via Adaptive Memory Management
School of Electronic and Computer Engineering, Peking University（电子与计算机工程学院，北京大学）；Department of Computer Science and Technology, Tsinghua University（计算机科学与技术系，清华大学）
Nanyang Technological University（南洋理工大学）；Institute for Infocomm Research（信息通信研究院）
Single modality action recognition on RGB or depth sequences has been extensively explored recently. It is generally accepted that each of these two modalities has different strengths and limitations for the task of action recognition. Therefore, analysis of the RGB+D videos can help us to better study the complementary properties of these two types of modalities and achieve higher levels of performance. In this paper, we propose a new deep autoencoder based shared-specific feature factorization network to separate input multimodal signals into a hierarchy of components. Further, based on the …
College of Information Technology, Gachon University（嘉泉大学信息技术学院）；Electronics and Telecommunications Research Institute (ETRI)（韩国电子通信研究院（ETRI））
Occlusion processing is a key issue in pedestrian attribute recognition (PAR). Nevertheless, several existing video-based PAR methods have not yet considered occlusion handling in depth. In this paper, we formulate finding non-occluded frames as sparsity-based temporal attention of a crowded video. In this manner, a model is guided not to pay attention to the occluded frame. However, temporal sparsity cannot include a correlation between attributes when occlusion occurs. For example, "boots" and "shoe color" cannot be recognized when the foot is invisible. To solve the uncorrelated attention i…
University of California（加利福尼亚大学）；Google（谷歌）；Seoul National University（首尔国立大学）
We propose a weakly supervised temporal action localization algorithm on untrimmed videos using convolutional neural networks. Our algorithm learns from video-level class labels and predicts temporal intervals of human actions with no requirement of temporal localization annotations. We design our network to identify a sparse subset of key segments associated with target actions in a video using an attention module and fuse the key segments through adaptive temporal pooling. Our loss function is comprised of two terms that minimize the video-level action classification error and enforce the sp…
T2SGrid: Temporal-to-Spatial Gridification for Video Temporal Grounding
South China University of Technology（南方科技大学）；Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ)（广东人工智能与数字经济实验室）；Bytedance Inc（字节跳动公司）
In this paper, we propose a novel sparse coding and counting method under Bayesian framwork for visual tracking. In contrast to existing methods, the proposed method employs the combination of L0 and L1 norm to regularize the linear coefficients of incrementally updated linear basis. The sparsity constraint enables the tracker to effectively handle difficult challenges, such as occlusion or image corruption. To achieve realtime processing, we propose a fast and efficient numerical algorithm for solving the proposed model. Although it is an NP-hard problem, the proposed accelerated proximal gra…
Attention Sparsity is Input-Stable: Training-Free Sparse Attention for Video Generation via Offline Sparsity Profiling and Online QK Co-Clustering
SKLCCSE, School of Computer Science and Engineering, Beihang University（软件学院，北京航空航天大学）；Beihang University（北京航空航天大学）；School of Computer Science, Peking University（北京大学计算机学院）；the State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences（多模态人工智能系统国家重点实验室，中国科学院自动化研究所）；School of Information Science and Technology, University of Science and Technology of China（信息科学与技术学院，中国科学技术大学）；BNRist, Tsinghua University（北京理工大学，清华大学）；Zhongguancun Academy（中关村学院）
Iowa State University（爱荷华州立大学）；Carnegie Mellon University（卡内基梅隆大学）
We study the problem of tracking (causally estimating) a time sequence of sparse spatial signals with changing sparsity patterns, as well as other unknown states, from a sequence of nonlinear observations corrupted by (possibly) non-Gaussian noise. In many applications, particularly those in visual tracking, the unknown state can be split into a small dimensional part, e.g. global motion, and a spatial signal, e.g. illumination or shape deformation. The spatial signal is often well modeled as being sparse in some domain. For a long sequence, its sparsity pattern can change over time, although …
Northeastern University（东北大学）；Roux Institute, Northeastern University（东北大学鲁克斯研究所）；University of Maine（缅因大学）；Khoury College of Computer Sciences, Northeastern University（东北大学胡里计算机科学学院）；Bouvé College of Health Sciences, Northeastern University（东北大学布维健康科学学院）
We present an end-to-end computer vision pipeline to detect non-nutritive sucking (NNS) -- an infant sucking pattern with no nutrition delivered -- as a potential biomarker for developmental delays, using off-the-shelf baby monitor video footage. One barrier to clinical (or algorithmic) assessment of NNS stems from its sparsity, requiring experts to wade through hours of footage to find minutes of relevant activity. Our NNS activity segmentation algorithm solves this problem by identifying periods of NNS with high certainty -- up to 94.0\% average precision and 84.9\% average recall across 30 …
RoboMirror: Understand Before You Imitate for Video to Humanoid Locomotion
Nanyang Technological University（南洋理工大学）；The Hong Kong University of Science and Technology（香港科技大学）；Beijing Academy of Artificial Intelligence（北京人工智能研究院）；Xi’an Jiaotong University（西安交通大学）；Chinese Academy of Sciences（中国科学院）；Shanghai Jiao Tong University（上海交通大学）；University of Sydney（悉尼大学）；Peking University（北京大学）
Outdoor videos sometimes contain unexpected rain streaks due to the rainy weather, which bring negative effects on subsequent computer vision applications, e.g., video surveillance, object recognition and tracking, etc. In this paper, we propose a directional regularized tensor-based video deraining model by taking into consideration the arbitrary direction of rain streaks. In particular, the sparsity of rain streaks in spatial and derivative domains, the spatiotemporal sparsity and low-rank property of video background are incorporated into the proposed method. Different from many previous me…
Tsinghua University（清华大学）；Shanghai AI Laboratory（上海人工智能实验室）；ETH Zürich（苏黎世联邦理工学院）；ByteDance Inc（字节跳动公司）；University of Texas at Dallas（德克萨斯大学达拉斯分校）
The high computational costs of video super-resolution (VSR) models hinder their deployment on resource-limited devices, (e.g., smartphones and drones). Existing VSR models contain considerable redundant filters, which drag down the inference efficiency. To prune these unimportant filters, we develop a structured pruning scheme called Structured Sparsity Learning (SSL) according to the properties of VSR. In SSL, we design pruning schemes for several key components in VSR models, including residual blocks, recurrent networks, and upsampling networks. Specifically, we develop a Residual Sparsity…
We present a video compressive sensing framework, termed kt-CSLDS, to accelerate the image acquisition process of dynamic magnetic resonance imaging (MRI). We are inspired by a state-of-the-art model for video compressive sensing that utilizes a linear dynamical system (LDS) to model the motion manifold. Given compressive measurements, the state sequence of an LDS can be first estimated using system identification techniques. We then reconstruct the observation matrix using a joint structured sparsity assumption. In particular, we minimize an objective function with a mixture of wavelet sparsi…
Rectified SpaAttn: Revisiting Attention Sparsity for Efficient Video Generation
Institute of Automation, Chinese Academy of Sciences（中国科学院自动化研究所）；School of Artificial Intelligence, University of Chinese Academy of Sciences（中国科学院大学人工智能学院）
DynamicRad: Content-Adaptive Sparse Attention for Long Video Diffusion
University of Electronic Science and Technology of China（电子科学与技术大学）；Shenzhen Institute for Advanced Study, UESTC（深圳先进研究院）；Computational Mathematics, Science, & Engineering at Michigan State University (MSU)（密歇根州立大学计算数学、科学与工程）
