27 Feb
13:30 Doctoral defense IC3 Auditorium
Reidentification of Vehicles and People: Methods and Applications
Edgar Rodolfo Quispe Condori
Advisor / Teacher
Helio Pedrini
Brief summary
Re-identification (ReID) is a problem in the field of computer vision that aims to combine instances of entities (for example, people, vehicles or luggage) through a system of cameras that do not overlap. Several factors make the task challenging, such as occlusions, lighting conditions, camera settings, different points of view and complex scene backgrounds. Different application domains can benefit from the ReID problem, for example surveillance and security, tracking, forensics and robotics. In this thesis, we investigate this task in a broad and comprehensive scheme. As research on ReID has evolved into a more realistic scenario, our research also follows this trend. We begin with the proposal of a supervised method for ReID of people that improves the representation of attributes of a neural network, learning discriminative information from regions of low activation. Next, we move to a scenario with more identifiers (IDs) and develop a method that efficiently leverages attribute labels for vehicle ReIDs. This method distills task-specific attribute information rather than following the literature that uses all attribute information. Finally, we apply the ReID to another task called Multi-Object Tracking. We investigated a less explored problem in the literature and showed that the adaptive use of the ReID attribute on highly occluded objects during training leads to better performance. We evaluated our three proposed methods on widely used datasets and showed that the results are competitive.
Examination Board
Hélio Pedrini IC / UNICAMP
Carlos Antônio Caetano Júnior Samsung
Ronaldo Cristiano Prati CMCC / UFABC
Alexandre Mello Ferreira FUNCAMP
André Santanchè IC / UNICAMP
Rafael de Oliveira Werneck IC / UNICAMP
Filipe de Oliveira Costa CPqD
Moacir Antonelli Ponti ICMC / USP