2014 ECCV, Ning Zhang, Jeff Donahue, Ross Girshick, Trevor Darrell.
Semantic part localization can facilitate fine-grained categorization
-
part-based RCNN:
- use RCNN to treat objects and parts as independent object categories
- train a one-versus-all linear SVM on feature descriptors extracted over region proposals
-
Geometric constraints:
- consider the locations of object of parts to choose best ones
-
Fine-grained categrization:
- Use fine-tuned CNN model to extract feature for object and parts independently
- one-versus-all linear SVM to classify