1st Workshop on
Target Re-Identification and
Multi-Target Multi-Camera Tracking
In conjunction with CVPR 2017
July 2017, Honolulu
This workshop brings together researchers from two subfields of computer vision that have seen growing activity in the past few years: Target Re-Identification (ReID) and Multi-Target Multi-Camera Tracking (MTMCT). This workshop explores opportunities for cross-fertilization between these two thriving and important subfields of computer vision and aims to deepen discussions on future directions.
This one day workshop will host invited speakers, poster sessions and oral presentations. We encourage authors to explore pollination between the fields of ReID and MTMCT and take on research questions, rather than just obtaining state-of-the-art results on benchmarks. Examples of such questions are: Call for papers
- How much do initial detections influence MTMCT or ReID?
- How and which ReID descriptors can be used by MTMCT systems?
- What can we learn by evaluating a MTMCT system in terms of ReID, and vice-versa?
- What makes these problems hard (and easy)?
- How can ReID and MTMCT benefit each other?
- How can MTMCT and ReID capitalize on recent large-scale datasets?
- How to deal with large-scale indexing/optimization in ReID and MTMCT?
- Do we need largely annotated datasets for ReID and MTMCT?
- Do semantic attributes help in matching identities in ReID and MTMCT?
- How can re-ranking schemes improve ReID performance?
Let a query be an image that tightly bounds a single instance of a target of interest (people, vehicles, etc.). Assume that a gallery contains images of several targets in this format as well as distractor images with arbitrary content. Problem definitions and motivationTarget re-identification retrieves all and only the gallery images of the same target as the query.
Given a set of videos taken by multiple cameras, multi-target multi-camera tracking places tight bounding boxes around all targets in the videos and partitions the boxes into sets called trajectories. A trajectory is the set of all boxes that bound a unique target, ordered by time.
These two problems are clearly different. However, they share several common aspects as well. Here, you can find some of the (dis)similarities that we could think of. Can you come up with more?
- The gallery in target re-identification may or may not be a set of videos and both targets and queries are assumed to have been isolated ahead of time. The input to multi-target multi-camera tracking on the other hand is necessarily a set of videos.
- Multi-target multi-camera tracking matches all targets symmetrically while target re-identification distinguishes between query and response explicitly.
- They assume a semantic notion of “identity” in that only ground truth can tell if the targets in two bounding boxes share the same identity.
- They could benefit from a detector that separates targets from non-targets.
- When the gallery in target re-identification is a set of videos, both problems can use similar pre-processing techniques. These might rely on knowing the camera topology, considering time elapsed between observations, modeling changes in illumination and viewpoint across cameras, etc.
- Both problems require annotated databases of videos or images and some databases can work for both problems.
- Some components of the solution to either problem can be used to solve the other.
To submit a new paper to the workshop, you have to do so through the CMT website. Coming soon!
Cordelia Schmid, INRIA
Xiaogang Wang, Chinese University of Hong Kong
Amit Roy-Chowdhury, University of California Riverside
Mubarak Shah, University of Central Florida
Chunhua Shen, University of Adelaide
Wei-Shi Zheng, Sun Yat-sen University
Anton Milan, University of Adelaide
Laura Leal-Taixé, Technical University Munich Title TBA
The following program will change according to the number of accepted papers and assigned oral presentations.
|9:00||Welcome and opening remarks|
Rita Cucchiara (University of Modena and Reggio Emilia)
Wen Gao (Peking University)
Shaogang Gong (Queen Mary University of London)
Thomas S. Huang (University of Illinois Urbana-Champaign)
Ergys Ristani (Duke University)
Francesco Solera (University of Modena and Reggio Emilia)
Qi Tian (University of Texas at San Antonio)
Carlo Tomasi (Duke University)
Program committee chairs:
Simone Calderara (University of Modena and Reggio Emilia)
Cees G.M. Snoek (University of Amsterdam)
Jingdong Wang (Microsoft Research)
Shiliang Zhang (Peking University)
Program committee members:
Horst Bischof, Roman Pflugfelder, Octavia Camps, Amit Roy-Chowdhury, Andrea Cavallaro, Peter Roth, Chen Change Loy, Ling Shao, Dimitrios Makris, Slawomir Bak, Ratnesh Kumar, Gerard Medioni, Xinchao Wang, Christian Micheloni, Liang Zheng and Vittorio Murino.
For any information, please send an Contacte-mail to Ergys Ristani, Francesco Solera and Jingdong Wang.