http://openaccess.thecvf.com/content_cvpr_2017/papers/Fu_Look_Closer_to_CVPR_2017_paper.pdf
PyTorch Implementation of RACNN
Does not replicate results
src/networks.py contain all networks implemented for the paper
src/manager... are training/eval managers for all networks
src/run... and src/main... are interfaces to run training
Run ./run_model_coords.sh 3 1 vgg 30 ../data/CUBS 1e-4 in src to train Attention Proposal Networks on best randomly generated subregions of interest.
Rename satisfactory trained APN to apn2.pt.pt in checkpoints
Run ./run_model.sh 3 1 vgg 30 ../data/CUBS 2 in src to train a two scale RACNN initialized with the APN trained above.
Rename satisfactory trained RACNN to racnn2.pt.pt in checkpoints
Run ./run_model_multi_scale.sh 3 1 vgg 30 ../data/CUBS 2 in srcto train a fully connected layer on a multiscale representation extracted using the RACNN trained above.