Evaluating Person Re-identification Performance on GAN-enhanced Datasets

Daniel Hofer, Wolfgang Ertel


Person re-identification remains a hard task for AI systems because high intra-class variance across different cameras, angles and lighting conditions make it difficult to create a reliable re-identification system. Since only small datasets for person re-id tasks are available, in recent years Generative Adversarial Networks (GANs) have become popular to improve intra-class variance to train more robust re-identification frameworks. In this work we evaluate an Inception-ResNet-v2 using triplet loss, introduced by (Weinberger and Saul, 2009), which works very well for face re-identification and use it for full-body person re-identification. The network is trained without GAN generated images to get a baseline accuracy of the network. In further experiments, the network is trained by adding constantly rising amounts of synthetic images produced by two image generators using different generating approaches.


Paper Citation