{"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/scene-flow-to-action-map-a-new-representation","title":"Scene Flow to Action Map: A New Representation for RGB-D based Action Recognition with Convolutional Neural Networks","arxiv_id":"1702.08652","date":"2017-02-28","proceeding":"CVPR 2017 7","authors":["Pichao Wang","Wanqing Li","Zhimin Gao","Yuyao Zhang","Chang Tang","Philip Ogunbona"],"abstract":"Scene flow describes the motion of 3D objects in real world and potentially\ncould be the basis of a good feature for 3D action recognition. However, its\nuse for action recognition, especially in the context of convolutional neural\nnetworks (ConvNets), has not been previously studied. In this paper, we propose\nthe extraction and use of scene flow for action recognition from RGB-D data.\nPrevious works have considered the depth and RGB modalities as separate\nchannels and extract features for later fusion. We take a different approach\nand consider the modalities as one entity, thus allowing feature extraction for\naction recognition at the beginning. Two key questions about the use of scene\nflow for action recognition are addressed: how to organize the scene flow\nvectors and how to represent the long term dynamics of videos based on scene\nflow. In order to calculate the scene flow correctly on the available datasets,\nwe propose an effective self-calibration method to align the RGB and depth data\nspatially without knowledge of the camera parameters. Based on the scene flow\nvectors, we propose a new representation, namely, Scene Flow to Action Map\n(SFAM), that describes several long term spatio-temporal dynamics for action\nrecognition. We adopt a channel transform kernel to transform the scene flow\nvectors to an optimal color space analogous to RGB. This transformation takes\nbetter advantage of the trained ConvNets models over ImageNet. Experimental\nresults indicate that this new representation can surpass the performance of\nstate-of-the-art methods on two large public datasets.","url_abs":"http://arxiv.org/abs/1702.08652v3","url_pdf":"http://arxiv.org/pdf/1702.08652v3.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-human-action-recognition","task_name":"3D Action Recognition"},{"task_slug":"action-recognition-in-videos","task_name":"Action Recognition"},{"task_slug":"action-recognition","task_name":"Temporal Action Localization"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/hand-gesture-recognition-on-chalearn-val","task":"Hand Gesture Recognition","dataset":"ChaLearn val","model":"Scene Flow","rank_in_archive_order":3,"of":3,"metrics":{"Accuracy":"36.27"},"uses_additional_data":false}],"syntology":{"syntology_url":"https://syntology.ai/paper/1702.08652","atlas_url":"https://app.syntology.ai/?focus=1702.08652","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}