Towards More Robust Person Re-identification for Smart Surveillance
NATIONAL ENGINEERING LABORATORY FOR VIDEO TECHNOLOGY
In this talk, I will introduce three of our latest CVPR'18 papers on person re-identification. Similarity learning is vital for person re-identification. Benefited from deep neural networks (DNN), current approaches can learn accurate similarity metrics and robust feature embeddings. However, most of them impose only local constraints for supervision. In the first part of my talk, I will introduce incorporating constraints on large image groups for similarity learning by combining the CRFs with deep neural networks. The proposed method aims to learn more robust visual similarity metrics for image pairs while taking into account the dependencies from all the images in the group. Extensive experiments demonstrate the effectiveness of our model that combines DNN and CRF for learning robust multi-scale local similarities. In the second part of my talk, I will introduce integrating a feature matching process into end-to-end learnable deep networks for handling person pose and view angle variations. Our proposed approach exploits Kronecker Product Matching to match corresponding features and conducts soft warping to align feature maps. The proposed approach outperforms state-of-the-art methods on public datasets. In the last part of my talk, I will introduce the background-bias problem in existing person re-ID datasets and algorithms. Existing deep learning models are biased to capture too much relevance between background appearances of person images. We identify this problem and proposed an online-background augmentation scheme and a person-region guided pooling deep neural network to solve this problem. Experiments demonstrate the robustness and effectiveness of our proposed algorithm.
Hongsheng Li will be an assistant professor in the Department of Electronic Engineering at the Chinese University of Hong Kong. He is currently a research assistant professor in the same lab. He received the bachelors degree in Automation from East China University of Science and Technology in 2006, and the doctorate degree in Computer Science from Lehigh University, United States in 2012. From 2013-2015, he was an associate professor in the School of Electronic Engineering at University of Electronic Science and Technology of China. He has published over 30 papers in top computer vision conferences, CVPR/ICCV/ECCV. He won the first place in Object Detection from Videos (VID) track of ImageNet challenge 2016 as the team leader and 2015 as a team co-leader. His research interests include computer vision, machine learning, and medical image analysis.