Personal tools
Laboratoire d'InfoRmatique en Image et Systèmes d'information

Skip to content. | Skip to navigation

Laboratoire d'InfoRmatique en Image et Systèmes d'information
UMR 5205 CNRS / INSA Lyon / Université Claude Bernard Lyon 1 / Université Lumière Lyon 2 / École Centrale de Lyon
You are here: Home > membres

Bonan Cuan


PhD student

Team Feature Extraction and Identification
Institution Institut National des Sciences Appliquées de Lyon
Location Jules Verne (INSA)
E-mail bonan.cuan at
Contact details Publications Thesis
Subject Tracking by Re-identification: A Region-based Correspondence Network
Abstract Object tracking is a classic and important problem in computer vision. As Yilmaz, Javed, and Shah [1] pointed, object tracking is a key step in video analysis and is pertinent in visual tasks like automated video surveillance, human-computer interaction, unmanned/self-driving vehicle system, video compression, etc.
After decades of research, object tracking remains a challenge. Occlusions, scene clutter, abrupt motion of object and/or camera, non-rigid object deformation, all such challenges prevent tracking algorithms from constantly and precisely finding object trajectories. The situation deteriorates in Multi-Object Tracking (MOT) problem, i.e. tracking multiple interacting objects at the same time in a single image sequence.
An increasingly popular solution is Tracking-by-Detection approach. Recent success of object detection algorithms, especially those based on deep learning, e.g. Faster R-CNN [2], provides a more robust way of locating potential objects. Given detected objects, tracking becomes a problem of finding the same object that occurs in different frames.
Re-identification (re-ID) is an approach of establishing correspondence among detected objects and thus appropriate for the Tracking-by-Detection task. Methods like metric learning [3] employ a Siamese neural network [4] to measure similarity or distance between pairs of object instances. With well-designed classification criteria, such clustering algorithms can serve as re-ID methods in tracking tasks.
In this thesis, we propose a hierarchical correspondence network to cope with tracking problem. As in Faster R-CNN, object detection is realized in two phases: a coarse but rapid Region Proposal Network (RPN) which examines region pyramids in an image, and a fine Fast R-CNN takes only the regions that may contain objects according to RPN as input. In our network, we introduce a third phase: the detected objects in different frames are then compared with each other in a deep Siamese net trained to re-ID all the instances of each object. All the three phases share the same deep feature map. The re-ID results are robust discriminative appearance model in Tracking-by-Detection schema like Multiple Hypothesis Tracking (MHT) [5].
Advisor Khalid Idrissi
Advisor Christophe Garcia

Last update : 2018-01-09 10:52:06