This repo contains code for the paper : Improving traffic sign recognition by active search
before doing anything be sure to either work in a virtual environement where the requirement.txt is install or in the docker image
-
first donwload and unzip all of the training, validation and test set from the mapilarry website for the traffic sign dataset
-
extract the patch
python thesis_data_search/patch_extraction.py --img_path=MAPILLARY_FOLDER/images --annot_path=MAPILLARY_FOLDER/annotations --output_path=YOUR__OUTPUT_FOLDER/patches- For the active loop on the 25 rarest traffic signs classes
python active_loop.py --top=50 --ep=5 --model=StandardNet --result_file=standard-net-al.pkl --path_data=path_to_data- For the active loop on synthetic data:
python active_loop.py --top=50 --ep=5 --limit_search=50 --model=StandardNet --result_file=non-synthetic-al.pkl --dataset=0
python active_loop.py --top=50 --ep=5 --limit_search=50 --model=StandardNet --result_file=synthetic-al.pkl --dataset=2- For the active loop with few shot learning:
python active_loop.py --top=50 --ep=5 --model=ProtoNetFull --result_file=proto-net-al.pkl- For the searching only:
python searching.py --model=StandardNet --result_file=standard-net-searching.pkl