{"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/semantic-alignment-finding-semantically","title":"Semantic Alignment: Finding Semantically Consistent Ground-truth for Facial Landmark Detection","arxiv_id":"1903.10661","date":"2019-03-26","proceeding":"CVPR 2019 6","authors":["Zhiwei Liu","Xiangyu Zhu","Guosheng Hu","Haiyun Guo","Ming Tang","Zhen Lei","Neil M. Robertson","Jinqiao Wang"],"abstract":"Recently, deep learning based facial landmark detection has achieved great\nsuccess. Despite this, we notice that the semantic ambiguity greatly degrades\nthe detection performance. Specifically, the semantic ambiguity means that some\nlandmarks (e.g. those evenly distributed along the face contour) do not have\nclear and accurate definition, causing inconsistent annotations by annotators.\nAccordingly, these inconsistent annotations, which are usually provided by\npublic databases, commonly work as the ground-truth to supervise network\ntraining, leading to the degraded accuracy. To our knowledge, little research\nhas investigated this problem. In this paper, we propose a novel probabilistic\nmodel which introduces a latent variable, i.e. the 'real' ground-truth which is\nsemantically consistent, to optimize. This framework couples two parts (1)\ntraining landmark detection CNN and (2) searching the 'real' ground-truth.\nThese two parts are alternatively optimized: the searched 'real' ground-truth\nsupervises the CNN training; and the trained CNN assists the searching of\n'real' ground-truth. In addition, to recover the unconfidently predicted\nlandmarks due to occlusion and low quality, we propose a global heatmap\ncorrection unit (GHCU) to correct outliers by considering the global face shape\nas a constraint. Extensive experiments on both image-based (300W and AFLW) and\nvideo-based (300-VW) databases demonstrate that our method effectively improves\nthe landmark detection accuracy and achieves the state of the art performance.","url_abs":"http://arxiv.org/abs/1903.10661v1","url_pdf":"http://arxiv.org/pdf/1903.10661v1.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":"face-alignment","task_name":"Face Alignment"},{"task_slug":"facial-landmark-detection","task_name":"Facial Landmark Detection"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/face-alignment-on-300w","task":"Face Alignment","dataset":"300W","model":"HGs + SA + Norm + GHCU","rank_in_archive_order":46,"of":48,"metrics":{"NME_inter-pupil (%, Challenge)":"6.38","NME_inter-pupil (%, Common)":"3.45","NME_inter-pupil (%, Full)":"4.02"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1903.10661","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}