HA-in-Fine-Grained-Classification

This repo includes the CUB-GHA (Gaze-based Human Attention) dataset and code of the paper "Human Attention in Fine-grained Classification" accepted to BMVC 2021.

CUB-GHA Dataset

To get the CUB-GHA (heatmap for each image) as shown in the paper, you can download from here (CUB-GHA.zip). Every image is saved under its index, and the index can be found in images.txt in CUB_200_2011.

If you would like to generate GHA by yourself. You need: (1) CUB-200-2011, which can be downloaded here. (2) some python packages: numpy, matplotlib, scipy, PIL, tqdm.

1. To generate the all fixation points in one heatmap for each image, as shown in the example below, please run the command: python generate_heatmap.py --CUB_dir_path <path_to_CUB> --CUB_GHA_save_path <path_to_save_CUB_GHA> --gaze_file_path ./Fixation.txt

   

2. To get single fixation heatmaps for each image, as shown in the example below, please run the command. Fixation belonging to one image will be saved under a directory named with its index. python generate_heatmap.py --single_fixation --CUB_dir_path <path_to_CUB> --CUB_GHA_save_path <path_to_save_CUB_GHA> --gaze_file_path ./Fixation.txt

   

   More settings can be found in the comments in the script. Please note that the fixation duration will not effect the fixation heatmaps in this mode.

3. Some comments of Fixation.txt:

   In "Fixation.txt", gaze data of each image in CUB can be found. Each line contains the following information:

   img_id, original_img_width, original_img_height, img_width_on_display, img_height_on_display, x_img_on_display, y_img_on_display, x_gaze_0, y_gaze_0, duration_gaze_0, x_gaze_1, y_gaze_1, duration_gaze_1, .... x_gaze_N, y_gaze_N, duration_gaze_N.

   "Fixation.txt" includes all gaze data from five runs of the data collection (after filtering gaze duration <0.1s). Inside the folder "data_5runs", you will find five files and each contains fixation in one run of the collection.

Experiment Code

In the folder CUB, you can find the code and instructions for experiments on CUB-200-2011.

In the folder CXR-Eye, you can find code and instructions for experiments on CXR-Eye.

If you use the CUB-GHA dataset or code in this repo in your research, please cite

@article{rong2021human,
title={Human Attention in Fine-grained Classification},
author={Rong, Yao and Xu, Wenjia and Akata, Zeynep and Kasneci, Enkelejda},
journal={arXiv preprint arXiv:2111.01628},
year={2021}
}

Contact me (yao.rong@uni-tuebingen.de) if you have any questions or suggestions.

We thank the following repos:

1. GazePointHeatMap for providing some functions of gaze visualization.

2. MMAL-Net for providing functions of training CUB.

3. cxr-eye-gaze for providing the dataset and functions of training on CXR-eye.