{"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/attention-is-all-we-need-nailing-down-object","title":"Attention is All We Need: Nailing Down Object-centric Attention for Egocentric Activity Recognition","arxiv_id":"1807.11794","date":"2018-07-31","proceeding":null,"authors":["Swathikiran Sudhakaran","Oswald Lanz"],"abstract":"In this paper we propose an end-to-end trainable deep neural network model\nfor egocentric activity recognition. Our model is built on the observation that\negocentric activities are highly characterized by the objects and their\nlocations in the video. Based on this, we develop a spatial attention mechanism\nthat enables the network to attend to regions containing objects that are\ncorrelated with the activity under consideration. We learn highly specialized\nattention maps for each frame using class-specific activations from a CNN\npre-trained for generic image recognition, and use them for spatio-temporal\nencoding of the video with a convolutional LSTM. Our model is trained in a\nweakly supervised setting using raw video-level activity-class labels.\nNonetheless, on standard egocentric activity benchmarks our model surpasses by\nup to +6% points recognition accuracy the currently best performing method that\nleverages hand segmentation and object location strong supervision for\ntraining. We visually analyze attention maps generated by the network,\nrevealing that the network successfully identifies the relevant objects present\nin the video frames which may explain the strong recognition performance. We\nalso discuss an extensive ablation analysis regarding the design choices.","url_abs":"http://arxiv.org/abs/1807.11794v1","url_pdf":"http://arxiv.org/pdf/1807.11794v1.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":[{"paper_slug":"attention-is-all-we-need-nailing-down-object","repo_url":"https://github.com/swathikirans/ego-rnn","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"activity-recognition","task_name":"Activity Recognition"},{"task_slug":"all","task_name":"All"},{"task_slug":"egocentric-activity-recognition","task_name":"Egocentric Activity Recognition"},{"task_slug":"hand-segmentation","task_name":"Hand Segmentation"}],"methods":[{"method_slug":"lstm","method_name":"LSTM"},{"method_slug":"sigmoid-activation","method_name":"Sigmoid Activation"},{"method_slug":"tanh-activation","method_name":"Tanh Activation"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/egocentric-activity-recognition-on-egtea-1","task":"Egocentric Activity Recognition","dataset":"EGTEA","model":"Ego-RNN","rank_in_archive_order":6,"of":6,"metrics":{"Average Accuracy":"60.8","Mean class accuracy":"-"},"uses_additional_data":false}],"syntology":{"syntology_url":null,"atlas_url":"https://app.syntology.ai/?focus=1807.11794","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}