Thanks https://github.com/timy90022/One-Shot-Object-Detection and https://github.com/jwyang/faster-rcnn.pytorch for providing the code which is used in this repository.

To setup repository, please follow these instructions:

Setup Faster R-CNN using instructions given in https://github.com/jwyang/faster-rcnn.pytorch
Download pre-trained weights for resnet-50 for image feature extractor and follow instructions given in image_model
Download pretrained weights for resnet-50 for sketch feature extractor and metadata: sketch_model
Store the downloaded data in ../data folder
Download the models from here

Download and Process QuickDraw

Download quickdraw dataset here and create the pickle file [processed_quick_draw_paths_common_classes.pkl] by running following commands:

python read_store_quick_draw.py
python get_train_test_quick_draw.py
python get_common_classes.py

Train and test the localization model

Enter default --sketch-path and --sketch_class_2_label paths in train and test files
To train - Run: python train_val_sketch_oneshot.py --dataset coco --net res50 --cuda --seen 7 --nw 10 --bs 10
To test - Run: python test_net_oneshot.py --dataset coco --net res50 --cuda --seen 7 --s <session> --checkepoch <epoch> --p <checkpoint>