Events
Your location now is: Home > Lectures > Content

Person Re-identification: Benchmarks and Our Solutions
Jul 13, 2017

窗体顶端

Title:

Person  Re-identification: Benchmarks and Our Solutions

Lecturer:

Prof. Qi Tian

Time:

2017-07-16  09:30:00

Venue:

Old  campus main building 3 District 430 meeting room

Lecturer  Profile

Qi Tian is currently  a Full Professor in the Department of Computer Science, the University of  Texas at San Antonio (UTSA). He was a tenured Associate Professor from  2008-2012 and a tenure-track Assistant Professor from 2002-2008. During  2008-2009, he took one-year Faculty Leave at Microsoft  Research Asia (MSRA) as Lead Researcher in the Media Computing Group.

Dr. Tian received  his Ph.D. in ECE from University of Illinois at Urbana-Champaign (UIUC) in  2002 and received his B.E. in Electronic Engineering from Tsinghua University  in 1992 and M.S. in ECE from Drexel University in 1996, respectively. Dr.  Tian’s research interests include multimedia information retrieval, computer  vision, pattern recognition and bioinformatics and published over 380  refereed journal and conference papers (including 90 IEEE/ACM Transactions  papers and 67 CCF Category A conference papers). He was the co-author of a  Best Paper in ACM ICMR 2015, a Best Paper in PCM 2013, a Best Paper in MMM  2013, a Best Paper in ACM ICIMCS 2012, a Top 10% Paper Award in MMSP 2011, a  Best Student Paper in ICASSP 2006, and co-author of a Best Student Paper  Candidate in ICME 2015, and a Best Paper Candidate in PCM 2007.

Dr. Tian research  projects are funded by ARO, NSF, DHS, Google, FXPAL, NEC, SALSI, CIAS, Akiira  Media Systems, HP, Blippar and UTSA. He received 2017 UTSA President’s  Distinguished Award for Research Achievement, 2016 UTSA Innovation Award,  2014 Research Achievement Awards from College of Science, UTSA, 2010 Google  Faculty Award, and 2010 ACM Service Award. He is the associate editor of IEEE  Transactions on Multimedia (TMM), IEEE Transactions on Circuits and Systems  for Video Technology (TCSVT), ACM Transactions on Multimedia Computing,  Communications, and Applications (TOMM), Multimedia System Journal (MMSJ),  and in the Editorial Board of Journal of Multimedia (JMM) and Journal of  Machine Vision and Applications (MVA).  Dr. Tian is the Guest Editor of  IEEE Transactions on Multimedia, Journal of Computer Vision and Image  Understanding, etc.

Dr. Tian is a Fellow  of IEEE. 田奇教授被评为2016年多媒体领域最有影响力的Top 10学者之一(by Aminer.org)。田奇教授也是教育部长江讲座教授和中科院海外评审专家,

URL:  http://www.cs.utsa.edu/~qitian

Email:  qi.tian@utsa.edu

Lecture Abstract

Person  re-identification (re-id) is a promising way towards automatic video  surveillance. As research hotspot in recent years, there has been an urgent  demand for building a solid benchmarking framework, including comprehensive  datasets and effective baselines.

To benchmark a large  scale person re-id dataset, we propose a new high quality frame-based dataset  for person re-identification titled “Market-1501”, which contains over 32,000  annotated bounding boxes, plus a distractor set of over 500K images. Different  from traditional datasets which use hand-drawn bounding boxes that are  unavailable under realistic settings, we produce the dataset with Deformable  Part Model (DPM) as pedestrian detector. Moreover, this dataset is collected  in an open system, where each identity has multiple images under each camera.  We propose an unsupervised Bag-of-Words representation and treat the person  re-identification as a special task of image search, which is demonstrated  very efficient and effective.

To further push the  person re-identification to practical applications, we propose a new video  based dataset titled “MARS”, which is the largest video re-id dataset to  date. Containing 1,261 identities and over 20,000 tracklets, it provides rich  visual information compared to image-based datasets. The tracklets are  automatically generated by the DPM as pedestrian detector and the GMMCP  tracker. Extensive evaluation of the state-of-the-art methods including the  space-time descriptors are presented. We further show that CNN in classification  mode can be trained from scratch using the consecutive bounding boxes of each  identity.

Finally, we present  “Person Re-identification in the Wild (PRW)” dataset for evaluating  end-to-end re-id methods from raw video frames to the identification results.  We address the performance of various combinations of detectors and  recognizers, mechanisms for pedestrian detection to help improve overall  re-identification accuracy and assessing the effectiveness of different  detectors for re-identification. A discriminatively trained ID-discriminative  Embedding (IDE) in the person subspace using convolutional neural network  (CNN) features and a Confidence Weighted Similarity (CWS) metric that  incorporates detection scores into similarity measurement are introduced to  aid the identification.

窗体底端

Previous:Ultrafast Laser Optics for Precision Metrology and Coherent Microscopy
Next:Impact of RAN centralization on 5G Access-Aggregation Networks

South Campus
Add: 266 Xinglong Section of Xifeng Road, Xi’an, Shaanxi 710126
Tel: 86-29-81891818
North Campus
Add: No. 2 South Taibai Road, Xi’an, Shaanxi 710071
Tel: 86-29-88202212