Learning Semantic Part Detectors from Google Images

In this talk I will present a novel method to train semantic part detectors without human annotations. Our framework learns rich models (appearance, location and viewpoint) entirely from Google Images, by collecting training instances for both parts and objects, and automatically connecting the two levels. It works incrementally, by learning from easy examples first, and then gradually adapting to harder ones. We evaluate our detectors on the challenging PASCAL-Part dataset. Results are encouraging and show potential in learning complex part models automatically.