Most transformer-based video encoders are limited to short temporal contexts due to their quadratic complexity. While various attempts have been made to extend this context, this has often come at the cost of both conceptual and computational complexity. We propose to instead re-purpose existing pre-trained video transformers by simply fine-tuning them to attend to memories derived non-parametrically from past activations. By leveraging redundancy reduction, our memory-consolidated vision transformer (MC-ViT) effortlessly extends its context far into the past and exhibits excellent scaling behavior when learning from longer videos. In doing so, MC-ViT sets a new state-of-the-art in long-context video understanding on EgoSchema, Perception Test, and Diving48, outperforming methods that benefit from orders of magnitude more parameters.
Recent incremental learning for action recognition usually stores representative videos to mitigate catastrophic forgetting. However, only a few bulky videos can be stored due to the limited memory. To address this problem, we propose FrameMaker, a memory-efficient video class-incremental learning approach that learns to produce a condensed frame for each selected video. Specifically, FrameMaker is mainly composed of two crucial components: Frame Condensing and Instance-Specific Prompt. The former is to reduce the memory cost by preserving only one condensed frame instead of the whole video, w…
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）
MovieChat: From Dense Token to Sparse Memory for Long Video Understanding
University of Washington（华盛顿大学）；Microsoft Research Asia（微软亚洲研究院）；Hong Kong University of Science and Technology (GZ)（香港科技大学（广州））；Donghua University（东华大学）；Zhejiang University（浙江大学）
In this paper, we propose a novel end-to-end trainable Video Question Answering (VideoQA) framework with three major components: 1) a new heterogeneous memory which can effectively learn global context information from appearance and motion features; 2) a redesigned question memory which helps understand the complex semantics of question and highlights queried subjects; and 3) a new multimodal fusion layer which performs multi-step reasoning by attending to relevant visual and textual hints with self-updated attention. Our VideoQA model firstly generates the global context-aware visual and tex…
AWS AI Labs（AWS人工智能实验室）；UNC Chapel Hill（北卡罗来纳大学教堂山分校）
We propose a memory efficient method, named Stochastic Backpropagation (SBP), for training deep neural networks on videos. It is based on the finding that gradients from incomplete execution for backpropagation can still effectively train the models with minimal accuracy loss, which attributes to the high redundancy of video. SBP keeps all forward paths but randomly and independently removes the backward paths for each network layer in each training step. It reduces the GPU memory cost by eliminating the need to cache activation values corresponding to the dropped backward paths, whose amount …
StreamMeCo: Long-Term Agent Memory Compression for Efficient Streaming Video Understanding
Shanghai Jiao Tong University（上海交通大学）；Fudan University（复旦大学）；Shanghai AI Laboratory（上海人工智能实验室）；Hong Kong University of Science and Technology（香港科学与技术大学）；MemTensor (Shanghai) Technology Co., Ltd.（MemTensor（上海）技术有限公司）
VideoEvolve: Co-Evolving Memory and Retrieval for Long Video Understanding
Alibaba Group（阿里巴巴集团）；University of Science and Technology of China（中国科学技术大学）；Shanghai Jiao Tong University（上海交通大学）；University of Chinese Academy of Sciences（中国科学院大学）；Beihang University（北京航空航天大学）
MemoryCard: Topic-Aware Multi-Modal Clue Compression for Long-Video Question Answering
School of Computer Science and Engineering, Northeastern University（东北大学计算机科学与工程学院）；Department of Computer Science and Technology, Tsinghua University（清华大学计算机科学与技术系）；Digital China Group（数字中国集团）
Scaling the Long Video Understanding of Multimodal Large Language Models via Visual Memory Mechanism
Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University（厦门大学多媒体可信感知与高效计算教育部重点实验室）；National University of Defense Technology（国防科技大学）
VideoTapestry: Query-Adaptive Memory Refinement for Multi-Agent Long-Video Understanding
University of Science and Technology of China（中国科学技术大学）；Jilin University（吉林大学）；Hangzhou Dianzi University（杭州电子科技大学）；Institute of Artificial Intelligence, Hefei Comprehensive National Science Center（合肥综合性国家科学中心人工智能研究院）
Association and Consolidation: Evolutionary Memory-Enhanced Incremental Multi-View Clustering
Institute of Information Science, Beijing Jiaotong University（北京交通大学信息科学学院）；Visual Intelligence + X International Cooperation Joint Laboratory of MOE（教育部视觉智能+X国际合作联合实验室）；School of Computer Science, Nanjing University of Posts and Telecommunications（南京邮电大学计算机科学学院）；School of Computer Science and Technology, Harbin Institute of Technology（哈尔滨工业大学计算机科学与技术学院）
Synaptic memory consolidation has been heralded as one of the key mechanisms for supporting continual learning in neuromorphic Artificial Intelligence (AI) systems. Here we report that a Fowler-Nordheim (FN) quantum-tunneling device can implement synaptic memory consolidation similar to what can be achieved by algorithmic consolidation models like the cascade and the elastic weight consolidation (EWC) models. The proposed FN-synapse not only stores the synaptic weight but also stores the synapse's historical usage statistic on the device itself. We also show that the operation of the FN-synaps…
Video-EM: Event-Centric Episodic Memory for Long-Form Video Understanding
City University of Hong Kong（香港城市大学）；University of Science and Technology of China（中国科学技术大学）；Tianjin University（天津大学）；Nanjing University（南京大学）；Zhejiang University（浙江大学）；The Institute of Artificial Intelligence (TeleAI), China Telecom（中国电信人工智能研究所（TeleAI））
We introduce a one-shot learning approach for video object tracking. The proposed algorithm requires seeing the object to be tracked only once, and employs an external memory to store and remember the evolving features of the foreground object as well as backgrounds over time during tracking. With the relevant memory retrieved and updated in each tracking, our tracking model is capable of maintaining long-term memory of the object, and thus can naturally deal with hard tracking scenarios including partial and total occlusion, motion changes and large scale and shape variations. In our experime…
Advanced Research Lab, NavInfo Europe（四维图新欧洲高级研究实验室）
Continual learning (CL) in the brain is facilitated by a complex set of mechanisms. This includes the interplay of multiple memory systems for consolidating information as posited by the complementary learning systems (CLS) theory and synaptic consolidation for protecting the acquired knowledge from erasure. Thus, we propose a general CL method that creates a synergy between SYNaptic consolidation and dual memory Experience Replay (SYNERgy). Our method maintains a semantic memory that accumulates and consolidates information across the tasks and interacts with the episodic memory for effective…
Barcelona Supercomputing Center (BSC)（巴塞罗那超级计算中心）；Universitat Politècnica de Catalunya（加泰罗尼亚理工大学）；Intel Iberia（英特尔伊比利亚）
As the number of installed cameras grows, so do the compute resources required to process and analyze all the images captured by these cameras. Video analytics enables new use cases, such as smart cities or autonomous driving. At the same time, it urges service providers to install additional compute resources to cope with the demand while the strict latency requirements push compute towards the end of the network, forming a geographically distributed and heterogeneous set of compute locations, shared and resource-constrained. Such landscape (shared and distributed locations) forces us to desi…
Semi-parametric Memory Consolidation: Towards Brain-like Deep Continual Learning
Centre for Artificial Intelligence and Robotics（人工智能与机器人中心）；Hong Kong Institute of Science and Innovation（香港科学与创新研究院）；Chinese Academy of Sciences（中国科学院）；Department of Computer Science（计算机科学系）；City University of Hong Kong（香港城市大学）；The Department of Neurosurgery（神经外科部门）；Peking University First Hospital（北京大学第一医院）；State Key Laboratory of Multimodal Artificial Intelligence Systems（多模态人工智能系统国家重点实验室）；Institute of Automation（自动化研究所）；Hong Kong SAR, China（香港特别行政区，中国）
University of Illinois Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校）
We present XMem, a video object segmentation architecture for long videos with unified feature memory stores inspired by the Atkinson-Shiffrin memory model. Prior work on video object segmentation typically only uses one type of feature memory. For videos longer than a minute, a single feature memory model tightly links memory consumption and accuracy. In contrast, following the Atkinson-Shiffrin model, we develop an architecture that incorporates multiple independent yet deeply-connected feature memory stores: a rapidly updated sensory memory, a high-resolution working memory, and a compact t…
The University of Tokyo（东京大学）；RIKEN（理化学研究所）；National Institutes of Health Clinical Center（美国国立卫生研究院临床中心）；eBay Inc.（易贝公司）；Temple University（天普大学）
Existing video domain adaption (DA) methods need to store all temporal combinations of video frames or pair the source and target videos, which are memory cost expensive and can't scale up to long videos. To address these limitations, we propose a memory-efficient graph-based video DA approach as follows. At first our method models each source or target video by a graph: nodes represent video frames and edges represent the temporal or visual similarity relationship between frames. We use a graph attention network to learn the weight of individual frames and simultaneously align the source and …
This paper addresses the task of segmenting moving objects in unconstrained videos. We introduce a novel two-stream neural network with an explicit memory module to achieve this. The two streams of the network encode spatial and temporal features in a video sequence respectively, while the memory module captures the evolution of objects over time. The module to build a "visual memory" in video, i.e., a joint representation of all the video frames, is realized with a convolutional recurrent unit learned from a small number of training video sequences. Given a video frame as input, our approach …
Tencent（腾讯）；Southern University of Science and Technology（南方科技大学）
Typical techniques for video captioning follow the encoder-decoder framework, which can only focus on one source video being processed. A potential disadvantage of such design is that it cannot capture the multiple visual context information of a word appearing in more than one relevant videos in training data. To tackle this limitation, we propose the Memory-Attended Recurrent Network (MARN) for video captioning, in which a memory structure is designed to explore the full-spectrum correspondence between a word and its various similar visual contexts across videos in training data. Thus, our m…
We introduce Spatial-Temporal Memory Networks for video object detection. At its core, a novel Spatial-Temporal Memory module (STMM) serves as the recurrent computation unit to model long-term temporal appearance and motion dynamics. The STMM's design enables full integration of pretrained backbone CNN weights, which we find to be critical for accurate detection. Furthermore, in order to tackle object motion in videos, we propose a novel MatchTrans module to align the spatial-temporal memory from frame to frame. Our method produces state-of-the-art results on the benchmark ImageNet VID dataset…
Video-based person re-identification (reID) aims to retrieve person videos with the same identity as a query person across multiple cameras. Spatial and temporal distractors in person videos, such as background clutter and partial occlusions over frames, respectively, make this task much more challenging than image-based person reID. We observe that spatial distractors appear consistently in a particular location, and temporal distractors show several patterns, e.g., partial occlusions occur in the first few frames, where such patterns provide informative cues for predicting which frames to fo…
Matching-based networks have achieved state-of-the-art performance for video object segmentation (VOS) tasks by storing every-k frames in an external memory bank for future inference. Storing the intermediate frames' predictions provides the network with richer cues for segmenting an object in the current frame. However, the size of the memory bank gradually increases with the length of the video, which slows down inference speed and makes it impractical to handle arbitrary length videos. This paper proposes an adaptive memory bank strategy for matching-based networks for semi-supervised video…
Center for Research on Intelligent Perception and Computing（智能感知与计算研究中心）；National Laboratory of Pattern Recognition（模式识别国家重点实验室）；Center for Excellence in Brain Science and Intelligence Technology（脑科学与智能技术卓越创新中心）；Institute of Automation, Chinese Academy of Sciences（中国科学院自动化研究所）
Video captioning which automatically translates video clips into natural language sentences is a very important task in computer vision. By virtue of recent deep learning technologies, e.g., convolutional neural networks (CNNs) and recurrent neural networks (RNNs), video captioning has made great progress. However, learning an effective mapping from visual sequence space to language space is still a challenging problem. In this paper, we propose a Multimodal Memory Model (M3) to describe videos, which builds a visual and textual shared memory to model the long-term visual-textual dependency an…
We address the problem of story-based temporal summarization of long 360° videos. We propose a novel memory network model named Past-Future Memory Network (PFMN), in which we first compute the scores of 81 normal field of view (NFOV) region proposals cropped from the input 360° video, and then recover a latent, collective summary using the network with two external memories that store the embeddings of previously selected subshots and future candidate subshots. Our major contributions are two-fold. First, our work is the first to address story-based temporal summarization of 360° videos. Secon…
Action anticipation and forecasting in videos do not require a hat-trick, as far as there are signs in the context to foresee how actions are going to be deployed. Capturing these signs is hard because the context includes the past. We propose an end-to-end network for action anticipation and forecasting with memory, to both anticipate the current action and foresee the next one. Experiments on action sequence datasets show excellent results indicating that training on histories with a dynamic memory can significantly improve forecasting performance.
