- structure code identity, pose
- appearance code: clothing, style, texture
generative module
self-identity training
cross-identity training
discirminative module
primary feature learning: teacher-student KL divergency
fine-grained feature mining
Person re-identify: make model robust to variants
novelty:
- end to end integrate discriminative and generative learning in a unified work.
- take advantage of the existing intra-dataset pose variants as well as other diversities beyond pose.