Training
A 4-step training algorithm is adopted:
- Finetune RPN using labeled image bounding boxes after initialized with ImageNet data
- Train the classification network (initialized with IMageNet data) by Fast RCNN using proposals generated from step 1 RPN
- Fix the shared convolutional layers, only fine-tune only the layers unique to RPN
- Fix the shared convolutional layers, only fine-tune the layers unique to Fast RCNN