Improving traffic sign recognition by active search

This repo contains code for the paper : "Learning to recognise rare traffic signs"

How to use this code

before doing anything be sure to either work in a virtual environement where the requirement.txt is install or in the docker image

Download the dataset and extract the patches

• 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 src/patch_extraction.py --img_path=MAPILLARY_FOLDER/images --annot_path=MAPILLARY_FOLDER/annotations --output_path=YOUR__OUTPUT_FOLDER/patches

run the script

• 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

• 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