We will present efficient and sc alable training framework that incorporates large- scale image-text pairs to achieve open-vocabulary object detection (OVD). Unlike previous OVD framew orks that typically rely on a pre-trained vision-l anguage model (e.g.\, CLIP) or exploit image-text pairs via a pseudo labelling process\, DetCLIP dir ectly learns the fine-grained word-region alignmen t from massive image-text pairs in an end-to-end m anner. We employ a maximum word-region similarity between region proposals and textual words to guid e the contrastive objective. To enable the model t o gain localization capability while learning broa d concepts\, DetCLIP is trained with a hybrid supe rvision from detection\, grounding and image-text pair data under a unified data formulation. By joi ntly training with an alternating scheme and adopt ing low-resolution input for image-text pairs\, De tCLIP exploits image-text pair data efficiently an d effectively.\n
\nBiographyDr Wei ZHang joined Huawei in 2012. Before that\, he was an assistant researcher in Shenzhen Institute of Advanced Technology Chinese Academy of Sciences and in The Chinese University of Hong Kong (CUHK) . He received his Ph.D. degree in computer science from CUHK in 2010\, his MS degree from Tsinghua U niversity in 2005 and his B.S. from Nankai Univers ity in 2002. Co-organizer of the “Self-supervised Learning for Next-Generation Industry-level Autono mous Driving” workshop at ECCV 2022 and ICCV 2021. \n LOCATION:EEDB Seminar Room\, Electrical Engineering and Onl ine (registration required) CONTACT:Dr Mark Leadbeater END:VEVENT END:VCALENDAR