Awesome

Gaze Target Estimation Inspired by Interactive Attention

This repository contains the code for our: Gaze Target Estimation inspired by interactive attention (VSG-IA)

Overview

The current repository has released the test code and pre-processed data on the GazeFollow dataset.

The complete code will be released as soon as possible.

• Evaluation code on GazeFollow/VideoTargetAttention datasets
• Pre-process code
• Training code

Instruction

1.Clone our repo and make directory "datasets".

git clone https://github.com/nkuhzx/VSG-IA
cd VSG-IA
mkdir datasets

2.Download the GazeFollow/VideoTargetAttention dataset refer to ejcgt's work and download them to datasets directory.

3.Download the preprocess file for test

sh download_preprocess.sh

4.Download the model weight for test

sh download_weight.sh

5.OPTIONAL SETTINGS

1)We provide a conda environment.yml file and you can re-create the environment we used.

conda env create -f environment.yml

2)We provide the test part of the gazefollow dataset for evaluation, you can download it to datasets directory and then unzip the gazefollow.zip in this directory.

3)You can download model weights for GazeFollow to your computer manually and model weights for VideoTargetAttention. The folder 'modelparas' need to contain the model_gazefollow.pt and model_videotargetattention.pt file.

A required structure is:

    VSG-IA/
    ├── datasets/
    │   ├── gazefollow/
    │   │   ├── test/
    │   │   ├── train/
    │   │   └── ...     
    │   ├── gazefollow_graphinfo/
    │   │   └── test_graph_data.hdf5
    │   ├── gazefollow_masks/
    │   │   └── test_masks/
    │   ├── gazefollow_annotation/
    │   │   ├── test_annotation.txt
    │   │   └── train_annotation.txt       
    │   ├── videotargetattention/
    │   │   ├── annotations/
    │   │   ├── images/
    │   │   └── ...     
    │   ├── videotargetattention_graphinfo/
    │   │   └── test_graph_data_vat.hdf5
    │   ├── videotargetattention_masks/
    │   │   ├── CBS This Morning/
    │   │   └── ...    
    │   ├── videotargetattention_annotation/
    │   │   ├── test_annotation.txt
    │   │   └── train_annotation.txt 
    ├── modelparas/
    │   ├── model_videotargetattention.txt
    │   └── model_gazefollow.gt
    ├── vsgia_model/
    │   ├── config/
    │   ├── dataset/
    │   ├── models/
    │   ├── utils/        
    │   ├── main.py
    │   ├── tester.py
    │   ├── main_vat.py
    │   ├── tester_vat.py
    │   └── __init__.py
    ├── download_dataset.sh
    ├── download_weight.sh
    ├── environment.yaml
    └── README.md
    

Evaluation

Evaluation on GazeFollow dataset

1.use gpu

cd vsgia_model
python main.py --gpu --is_test (use gpu)

2.only use cpu

cd vsgia_model
python main.py --is_test (only use cpu)

After the evaluation process, the program will reports the Average Distance, Minimum Distance and AUC.

Evaluation on VideoTargetAttention dataset

1.use gpu

cd vsgia_model
python main_vat.py --gpu --is_test (use gpu)

2.only use cpu

cd vsgia_model
python main_vat.py --is_test (only use cpu)

Acknowledgement

We thank the associate editors and reviewers for their constructive suggestions.