Training on Mass Surveillance Datasets Provides No Empirical Benefits

Abstract

Mass surveillance technologies rely on person re-identification (ReID), the task of matching images of the same person by appearance across changes in pose, viewpoint, time, and camera. ReID datasets are built from footage of pedestrians, students, and shoppers, recorded by surveillance cameras in public space without their consent. These datasets have drawn sustained ethical criticism, yet they remain openly hosted and widely used, defended by the claim that the models trained on them are useful for surveillance deployment. We test that claim on the tasks the footage was collected to enable, and it does not hold. We find that models finetuned on ReID datasets do not deliver their claimed deployment benefits: retrieval accuracy degrades outside the training corpus, tracking does not improve over a general pretrained encoder across eight architectures and three pedestrian-tracking benchmarks, and having more data does not change this. We further find that the representations this training produces encode stereotypes: clothing colour becomes far more decodable after finetuning, finetuned encoders are less likely to distinguish between different people wearing same clothing and saliency for the matching decision localises on clothing. We add empirical depth to the ethical case against ReID datasets, which until now made on the consent violation alone. We conclude that these datasets are built from non-consensual data and have no significant benefit in downstream deployment, thus these datasets should not be built.