Abstract:
Unsupervised domain adaptation severely suffers from huge domain gap in fine grained recognition tasks such as vehicle re-identification (re-id). Existing works either focus on fully unsupervised methods using tracklet or classical clustering based progressive learning techniques. However, these methods do not leverage the available large datasets which are labelled. Though this can be done using existing unsupervised domain adaptation techniques, the fine grained nature of vehicle re-id precludes from obtaining a sound performance. To this end, we propose a joint learning framework which disentangles ID and non-ID features and enforce the adaptation module to focus on the ID features only. Our model performs the following three steps; (i) the model encodes cross domain images into shared ID space and domain specific non-ID space, (ii) adaptation is performed using adversarial domain alignment and pseudo-label generation, and (iii) meta learning is applied to obtain better generalization. We perform experiments on AI CITY, VRIC and VeRI-776, and compare against various unsupervised techniques to show the efficacy of our model.
