Pre-trained RPN on the videos as describe in the paper can be found at link
Model was trained using the code based on Faster-RCNN (https://github.com/jwyang/faster-rcnn.pytorch). We use pseudo ground-truth boxes obtained using motion segments as described in the paper. To obtain proposals use the test_net.py.