Revisiting knowledge transfer for training object class detectors

We propose to revisit knowledge transfer for training object detectors on target classes with only weakly supervised training images. We present a unified knowledge transfer framework based on training a single neural network multi-class object detector over all source classes, organized in a semantic hierarchy. This provides proposal scoring functions at multiple levels in the hierarchy, which we use to guide object localization in the target training set. Compared to works using a manually engineered class-generic objectness measure as a vehicle for transfer, our learned top-level scoring function for 'entity' is much stronger. Compared to works that perform class-specific transfer from a few most related source classes to the target class, our framework enables to explore a broad rage of generality of transfer. Experiments on 200 object classes in the ILSVRC 2013 dataset show that our technique (1) leads to much greater performance improvements than manually engineered objectness; (2) outperforms the best reported transfer learning results on this dataset by a wide margin (+40% correct localization on the target training set, and +14% mAP on the target test set).