{"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/beyond-skip-connections-top-down-modulation","title":"Beyond Skip Connections: Top-Down Modulation for Object Detection","arxiv_id":"1612.06851","date":"2016-12-20","proceeding":null,"authors":["Abhinav Shrivastava","Rahul Sukthankar","Jitendra Malik","Abhinav Gupta"],"abstract":"In recent years, we have seen tremendous progress in the field of object\ndetection. Most of the recent improvements have been achieved by targeting\ndeeper feedforward networks. However, many hard object categories such as\nbottle, remote, etc. require representation of fine details and not just\ncoarse, semantic representations. But most of these fine details are lost in\nthe early convolutional layers. What we need is a way to incorporate finer\ndetails from lower layers into the detection architecture. Skip connections\nhave been proposed to combine high-level and low-level features, but we argue\nthat selecting the right features from low-level requires top-down contextual\ninformation. Inspired by the human visual pathway, in this paper we propose\ntop-down modulations as a way to incorporate fine details into the detection\nframework. Our approach supplements the standard bottom-up, feedforward ConvNet\nwith a top-down modulation (TDM) network, connected using lateral connections.\nThese connections are responsible for the modulation of lower layer filters,\nand the top-down network handles the selection and integration of contextual\ninformation and low-level features. The proposed TDM architecture provides a\nsignificant boost on the COCO testdev benchmark, achieving 28.6 AP for VGG16,\n35.2 AP for ResNet101, and 37.3 for InceptionResNetv2 network, without any\nbells and whistles (e.g., multi-scale, iterative box refinement, etc.).","url_abs":"http://arxiv.org/abs/1612.06851v2","url_pdf":"http://arxiv.org/pdf/1612.06851v2.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":"beyond-skip-connections-top-down-modulation","repo_url":"https://github.com/MTCloudVision/mxnet-dssd","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"mxnet","reach":null}],"tasks":[{"task_slug":"object","task_name":"Object"},{"task_slug":"object-detection","task_name":"Object Detection"},{"task_slug":"object-detection-1","task_name":"object-detection"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/object-detection-on-coco","task":"Object Detection","dataset":"COCO test-dev","model":"Faster R-CNN + TDM","rank_in_archive_order":221,"of":225,"metrics":{"box mAP":"36.8"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1612.06851","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}