{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/a2-nets-double-attention-networks","title":"$A^2$-Nets: Double Attention Networks","arxiv_id":"1810.11579","date":"2018-10-27","proceeding":null,"authors":["Yunpeng Chen","Yannis Kalantidis","Jianshu Li","Shuicheng Yan","Jiashi Feng"],"abstract":"Learning to capture long-range relations is fundamental to image/video\nrecognition. Existing CNN models generally rely on increasing depth to model\nsuch relations which is highly inefficient. In this work, we propose the\n\"double attention block\", a novel component that aggregates and propagates\ninformative global features from the entire spatio-temporal space of input\nimages/videos, enabling subsequent convolution layers to access features from\nthe entire space efficiently. The component is designed with a double attention\nmechanism in two steps, where the first step gathers features from the entire\nspace into a compact set through second-order attention pooling and the second\nstep adaptively selects and distributes features to each location via another\nattention. The proposed double attention block is easy to adopt and can be\nplugged into existing deep neural networks conveniently. We conduct extensive\nablation studies and experiments on both image and video recognition tasks for\nevaluating its performance. On the image recognition task, a ResNet-50 equipped\nwith our double attention blocks outperforms a much larger ResNet-152\narchitecture on ImageNet-1k dataset with over 40% less the number of parameters\nand less FLOPs. On the action recognition task, our proposed model achieves the\nstate-of-the-art results on the Kinetics and UCF-101 datasets with\nsignificantly higher efficiency than recent works.","url_abs":"http://arxiv.org/abs/1810.11579v1","url_pdf":"http://arxiv.org/pdf/1810.11579v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[],"tasks":[{"task_slug":"3d-absolute-human-pose-estimation","task_name":"3D Absolute Human Pose Estimation"},{"task_slug":"action-classification","task_name":"Action Classification"},{"task_slug":"action-recognition-in-videos","task_name":"Action Recognition"},{"task_slug":"action-recognition","task_name":"Temporal Action Localization"},{"task_slug":"video-recognition","task_name":"Video Recognition"}],"methods":[{"method_slug":"convolution","method_name":"Convolution"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/action-classification-on-kinetics-400","task":"Action Classification","dataset":"Kinetics-400","model":"A2 Net","rank_in_archive_order":162,"of":207,"metrics":{"Acc@1":"74.6","Acc@5":"91.5"},"uses_additional_data":false},{"leaderboard":"/sota/action-recognition-in-videos-on-ucf101","task":"Action Recognition","dataset":"UCF101","model":"A2-Net (ResNet-50)","rank_in_archive_order":38,"of":91,"metrics":{"3-fold Accuracy":"96.4"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1810.11579","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}