Train/Test on the same domain
Part-level features are effective. pose and 3D information are beneficial
**Smothness in feature space is beneficial **
Unsupervised domain adaptation
Evolution in state-of-the art Performance
image-image translation benefits UDA; identity-preserving property benefits further
mining discriminative cues in the target domain improves the UDA accuracy
挖掘 区别性大的线索 提高 unsupervised domain adaptation accuracy.
Video re-id: frame-level weights are important spatiotemporal attention network
feature memory modeling is critical for end to end person search
separate person detector and re-identification seem to be a better choice.
person re-id originates from person tracking; After years of study, person re-id features are improving tracking accuracies.
learning from synthetic data for person re-id **Synthetic training data can help to initialize deep networds **
the diversity of synthetic data can help improve the generalization performance of re-ID models
Future research questions
- is large-scale re-id really solved? what the performance will be if we scale up the gallery to 1 million or 10 million images? How to accelerate?
- In unsupervised domain adaptation, person re-id is a open-set problem. what is its relationship with UDA in image classification?
- What is the relationship between person detection and re-id? is it optimal to use re-id features in tracking?
- How do environmental factors affect re-id?