Person re-identification: Attribute-based feature evaluation
Chirine Riachy, Ahmed Bouridane · 2018
Matching people across various camera views at different times and locations is called person re-identification (re-id). Despite remarkable recent advances in the field, the problem is still particularly challenging due to significant viewpoint angle variations, illumination changes, background clutter, occlusions, motion blur and pose variations. Although these challenges are widely acknowledged within the re-id community, no previous work has attempted to analyze the extent to which they individually contribute to performance deterioration, and whether current feature representations have succeeded to mitigate their effect. To address this matter, 4 publically available single-shot datasets were fully and manually annotated for these attributes. Subsequently, 6 state-of-the-art feature representations were evaluated considering each attribute separately. Extensive experimentation using XQDA distance metric has clearly shown that illumination variation remains the major problem hindering accurate re-id. The results also provide insights towards the design of new feature representations to improve the performance.