主讲人：李建国研究员，Intel 中国研究院时间: 1 月 30 日（周五），下午 1:30-2:30地点: 复旦大学张江校区计算机楼 405联系人：姜育刚 ygj@fudan.edu.cn
Abstract: Deep learning makes a lot of breakthrough in computer vision tasks, and has been introduced in image retrieval to obtain some promising results. However, those methods only pick the output of some internal layers (FC or CONV) from CNN as holistic feature representation. This representation suffers from high dimension, and does not provide strategy to integrate context information for accuracy improvement. We introduce regional neural codes to exploit possible context information. Experiments show that the approach can obtain state-of-the-art accuracy on several benchmarks. We also utilize the sparse property of the neural codes, and design a data structure for efficient query processing, and compress the representation with small codes or tiny codes for storage saving. Nicely, experiments show that the small codes/tiny codes only decrease the accuracy slightly. I will also talk about possible directions to further improve accuracy.Bio: Jianguo is a staff research scientist and technique leader with Intel Labs China. He joined Intel Labs China in July 2006 after he got his PhD degree from Dept. Automation, Tsinghua university. His research focused on computer vision, large-scale machine learning and its applications in real life. He has published 30+ peer reviewed top-tier conferences and journal papers, including ICML, CVPR, IJCAI, MM, Micro, etc. His researches has made big impact/transferring across several Intel real products, including RealSense SDK, Pocket Avatar, Intel CPU hardware features, and so on.
