{"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/bottom-up-and-top-down-reasoning-with","title":"Bottom-Up and Top-Down Reasoning with Hierarchical Rectified Gaussians","arxiv_id":"1507.05699","date":"2015-07-21","proceeding":"CVPR 2016 6","authors":["Peiyun Hu","Deva Ramanan"],"abstract":"Convolutional neural nets (CNNs) have demonstrated remarkable performance in\nrecent history. Such approaches tend to work in a unidirectional bottom-up\nfeed-forward fashion. However, practical experience and biological evidence\ntells us that feedback plays a crucial role, particularly for detailed spatial\nunderstanding tasks. This work explores bidirectional architectures that also\nreason with top-down feedback: neural units are influenced by both lower and\nhigher-level units.\n  We do so by treating units as rectified latent variables in a quadratic\nenergy function, which can be seen as a hierarchical Rectified Gaussian model\n(RGs). We show that RGs can be optimized with a quadratic program (QP), that\ncan in turn be optimized with a recurrent neural network (with rectified linear\nunits). This allows RGs to be trained with GPU-optimized gradient descent. From\na theoretical perspective, RGs help establish a connection between CNNs and\nhierarchical probabilistic models. From a practical perspective, RGs are well\nsuited for detailed spatial tasks that can benefit from top-down reasoning. We\nillustrate them on the challenging task of keypoint localization under\nocclusions, where local bottom-up evidence may be misleading. We demonstrate\nstate-of-the-art results on challenging benchmarks.","url_abs":"http://arxiv.org/abs/1507.05699v5","url_pdf":"http://arxiv.org/pdf/1507.05699v5.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":"bottom-up-and-top-down-reasoning-with","repo_url":"https://github.com/peiyunh/rg-mpii","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"none","reach":{"status":"ok"}}],"tasks":[{"task_slug":null,"task_name":"GPU"},{"task_slug":"pose-estimation","task_name":"Pose Estimation"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/pose-estimation-on-mpii-human-pose","task":"Pose Estimation","dataset":"MPII Human Pose","model":"QP2","rank_in_archive_order":41,"of":46,"metrics":{"PCKh-0.5":"82.4"},"uses_additional_data":false},{"leaderboard":"/sota/pose-estimation-on-mpii-human-pose","task":"Pose Estimation","dataset":"MPII Human Pose","model":"QP1","rank_in_archive_order":45,"of":46,"metrics":{"PCKh-0.5":"81.1"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1507.05699","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}