visualize_attention

Visualize Attention Result

• This folder contains script to visual attention models result for each image.

• Before Runing scripts in this folder, make sure you have done operations listed in the Main README.md. You don't need to download pretrained ResNet-101 model and don't need to run script:

cd $REPO_ROOT
./scripts/generate_baseline.py

• Operations to create attention images are as follows:

1. Create attention info files for each image:

cd $REPO_ROOT/visualize_attention
python create_image_attention_info.py

This script will create one json file for each image in a newly create dir named debug_att_dir.

2. Merge attention info files into one json list:

python merge_image_attention_infos.py

This script will create one json file named merged_att_result.json in current dir.

3. Create images with attention bounding boxes:

python create_attention_images.py \
	--id2info_json merged_att_result.json \
	--image_dir 'your coco image root dir' \
	--top_n 3 \
	--result_dir 'choose one dir to save the attention images'

image_dir should be the parent dir of val2014 or train2014.

4. Result attention image:

Open result_dir dir to see attention images. One like this: