周晓巍
报告题目:Detector-free Feature Matching and its Applications in 3D Vision
个人简介:周晓巍,浙江大学“百人计划”研究员,国家级青年人才项目入选者。研究方向主要为三维视觉及其在混合现实、机器人等领域的应用。代表性工作包括NeuralBody, NeuralRecon, DeepSnake, PVNet, LoFTR等,十余次获得视觉领域顶级会议口头报告,多次入选CVPR最佳论文候选。曾获得浙江省自然科学一等奖,陆增镛CAD&CG高科技奖一等奖,CCF优秀图形开源贡献奖,入选斯坦福大学发布的2021全球前2%顶尖科学家榜单。担任国际顶级期刊IJCV编委、顶级会议CVPR/ICCV领域主席,图形学与混合现实研讨会(GAMES)执行委员会主席,视觉与学习研讨会(VALSE)常务委员,CSIG三维视觉专委会常务委员。
报告摘要:Local feature matching is a cornerstone for many 3D geometric vision problems such as structure from motion (SfM) and visual localization. Existing SfM and localization systems are mostly built upon detector-based feature matchers that rely on the successful detection of repeatable feature points across multiple views as the first step, which is difficult for texture-poor scenes. In this talk, I will introduce LoFTR, a detector-free approach to local feature matching and how it can be applied to SfM and visual localization, as well as our solution that won the image matching challenge (IMC) 2023.