As a junior, I worked as a Research Intern at the Computer Vision Lab at École Polytechnique Fédérale de Lausanne (EPFL). This was a part of NTU-EPFL’s Research Exchange program, and I was offered a research scholarship by EPFL for it. My supervisors were Dr. Horesh Ben Shitrit and Prof. Pascal Fua.
The aim of this project was to apply a person re-identification algorithm which uses trivial appearance cues like color and texture, and achieve satisfactory performance without training, on various datasets being used by the lab. Firstly, a literature review was done and the tracking results were visualized. After this, we explored three methods – (i) dominant colors, (ii) color histograms and (iii) color invariants for person re-identification. We applied these methods on two sports datasets – a volleyball sequence and a soccer sequence, and on one pedestrian dataset – a video shot in the lab at EPFL. Based on the accuracies calculated, we concluded that the signatures used in the color invariants method produce the best results, since they are parts-based signatures or signatures which take spacial information into account. The dominant colors and color histograms methods do not work very well, since they are holistic approaches. Usually, the color histograms approach gives better results than the dominant colors approach, however, dominant colors can work better in a situation where the illumination changes are not much and consistent dominant colors can be obtained to describe a person or team, like in the soccer dataset. Person re-identification based on appearance is a challenging problem, which works better if the clothes worn by the different people to be re-identified are very different in color and/or texture. In future, it would be interesting to see if superpixels can be used to divide the person into meaningful parts which can then be matched for person re-identification. Here is the technical report: http://prernac.com/reports/person_reid_chikersal_2014.pdf