Person Re-identification with Bias-controlled Adversarial Training
Sara Iodice, Krystian Mikolajczyk
Abstract
Inspired by the effectiveness of adversarial training in the area of Generative Adversarial Networks we present a new approach for learning feature representations in person re-identification. We investigate different types of bias that typically occur in re-ID scenarios, i.e., pose, body part and camera view, and propose a general approach to address them. We introduce an adversarial strategy for controlling bias, named Bias-controlled Adversarial framework (BCA), with two complementary branches to reduce or to enhance bias-related features. The results and comparison to the state of the art on different benchmarks show that our framework is an effective strategy for person re-identification. The performance improvements are in both full and partial views of persons.
Results
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Person Re-Identification | Market-1501 | Rank-1 | 93.1 | Bias-controlled Adversarial Training |
| Person Re-Identification | Market-1501 | mAP | 89.3 | Bias-controlled Adversarial Training |
| Person Re-Identification | DukeMTMC-reID | Rank-1 | 85.2 | Bias-controlled Adversarial Training |
| Person Re-Identification | DukeMTMC-reID | mAP | 74.8 | Bias-controlled Adversarial Training |