{"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/searching-for-efficient-multi-scale","title":"Searching for Efficient Multi-Scale Architectures for Dense Image Prediction","arxiv_id":"1809.04184","date":"2018-09-11","proceeding":"NeurIPS 2018 12","authors":["Liang-Chieh Chen","Maxwell D. Collins","Yukun Zhu","George Papandreou","Barret Zoph","Florian Schroff","Hartwig Adam","Jonathon Shlens"],"abstract":"The design of neural network architectures is an important component for\nachieving state-of-the-art performance with machine learning systems across a\nbroad array of tasks. Much work has endeavored to design and build\narchitectures automatically through clever construction of a search space\npaired with simple learning algorithms. Recent progress has demonstrated that\nsuch meta-learning methods may exceed scalable human-invented architectures on\nimage classification tasks. An open question is the degree to which such\nmethods may generalize to new domains. In this work we explore the construction\nof meta-learning techniques for dense image prediction focused on the tasks of\nscene parsing, person-part segmentation, and semantic image segmentation.\nConstructing viable search spaces in this domain is challenging because of the\nmulti-scale representation of visual information and the necessity to operate\non high resolution imagery. Based on a survey of techniques in dense image\nprediction, we construct a recursive search space and demonstrate that even\nwith efficient random search, we can identify architectures that outperform\nhuman-invented architectures and achieve state-of-the-art performance on three\ndense prediction tasks including 82.7\\% on Cityscapes (street scene parsing),\n71.3\\% on PASCAL-Person-Part (person-part segmentation), and 87.9\\% on PASCAL\nVOC 2012 (semantic image segmentation). Additionally, the resulting\narchitecture is more computationally efficient, requiring half the parameters\nand half the computational cost as previous state of the art systems.","url_abs":"http://arxiv.org/abs/1809.04184v1","url_pdf":"http://arxiv.org/pdf/1809.04184v1.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":"searching-for-efficient-multi-scale","repo_url":"https://github.com/tensorflow/models","is_official":0,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"tf","reach":null}],"tasks":[{"task_slug":"image-classification","task_name":"Image Classification"},{"task_slug":"image-segmentation","task_name":"Image Segmentation"},{"task_slug":"meta-learning","task_name":"Meta-Learning"},{"task_slug":"open-question","task_name":"Open-Ended Question Answering"},{"task_slug":"scene-parsing","task_name":"Scene Parsing"},{"task_slug":"segmentation","task_name":"Segmentation"},{"task_slug":"semantic-segmentation","task_name":"Semantic Segmentation"},{"task_slug":"street-scene-parsing","task_name":"Street Scene Parsing"},{"task_slug":"image-classification","task_name":"image-classification"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/human-part-segmentation-on-pascal-person-part-1","task":"Human Part Segmentation","dataset":"PASCAL-Person-Part","model":"DPC","rank_in_archive_order":1,"of":1,"metrics":{"mIoU":"71.34"},"uses_additional_data":false},{"leaderboard":"/sota/semantic-segmentation-on-cityscapes","task":"Semantic Segmentation","dataset":"Cityscapes test","model":"Dense Prediction Cell","rank_in_archive_order":25,"of":105,"metrics":{"Mean IoU (class)":"82.7%"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1809.04184","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}