14 May 2026, by Viktoria Wrobel

Photo: base.camp

Surveillance systems that match and track people across cameras rely on appearance encoders, which are commonly trained on Person Re-Identification (ReID) datasets. These datasets are ethically problematic: they contain images of people captured without meaningful consent and have been used by organizations conducting mass surveillance. Despite these concerns, their practical benefit over general-purpose encoders remains assumed rather than empirically tested. This project investigates whether ReID-finetuned encoders actually improve multi-object tracking performance compared to encoders trained with identical methods on non-surveillance data. Through controlled tracking experiments and probing analysis of learned representations, we test the hypothesis that ReID encoders learn dataset-specific biases rather than robust identity features. If ReID finetuning provides no measurable benefit, the privacy violations inherent in surveillance dataset collection lack empirical justification.