Methods for data-related problems in person re-ID

Abstract

In the last years, the ever-increasing need for public security has attracted wide attention in person re-ID. State-of-the-art techniques have achieved impressive results on academic datasets, which are nearly saturated. However, when it comes to deploying a re-ID system in a practical surveillance scenario, several challenges arise. 1) Full person views are often unavailable, and missing body parts make the comparison very challenging due to significant misalignment of the views. 2) Low diversity in training data introduces bias in re-ID systems. 3) The available data might come from different modalities, e.g., text and images. This thesis proposes Partial Matching Net (PMN) that detects body joints, aligns partial views, and hallucinates the missing parts based on the information present in the frame and a learned model of a person. The aligned and reconstructed views are then combined into a joint representation and used for matching images. The thesis also investigates different types of bias that typically occur in re-ID scenarios when the similarity between two persons is due to the same pose, body part, or camera view, rather than to the ID-related cues. It proposes a general approach to mitigate these effects named Bias-Control (BC) framework with two training streams leveraging adversarial and multitask learning to reduce bias-related features. Finally, the thesis investigates a novel mechanism for matching data across visual and text modalities. It proposes a framework Text (TAVD) with two complementary modules: Text attribute feature aggregation (TA) that aggregates multiple semantic attributes in a bimodal space for globally matching text descriptions with images and Visual feature decomposition (VD) which performs feature embedding for locally matching image regions with text attributes. The results and comparison to state of the art on different benchmarks show that the proposed solutions are effective strategies for person re-ID.