This code contains the training and test of Faster-RCNN

• Train
• Test
• Result

• Training process is implemented in train_frcnn.py
• You need to realize your own data_loader in roidb. deepdrive_data_loader.py is implemented for DeepDrive of AIChallenge dataset. You can also change it into VOC data loader or COCO data loader by overwriting get_minibatch()
• More details in train_frcnn.py. The default number of epochs is 50, and the learning rate decays by 0.1 at epoch 30. You can change it in train()
• The first 100 images containing the foreground anchors are visualized iin 'vis_anchors'. This is a test for generating anchors.

• Test is implemented in inference.py
• Put the testing image in to folder 'images' and modify inference.py
• You can visualize proposals output by RPN by uncommenting draw_proposals at line 181

• Detection
• Proposals