Ev-Eye: Event-based Eye Tracking with Motion Consistency

AIS 2024 Event-based Eye Tracking (CVPR Workshop 2024) 2nd Solution.

Environment

• Ubuntu22.04
• Python3.9
• Nvidia GPU1080Ti

Using the following command line for the installation of dependencies rquired to trn Ev-Eye.

git clone https://github.com/MrFisher97/Solution_For_Event_based_Eye_Tracking_Challenge.git
cd Solution_For_Event_based_Eye_Tracking_Challenge
conda create -n EvEye python==3.9
conda activate EvEye
conda install pytorch=1.11 torchvision cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt

Pretrained Model

• Google Drive: https://drive.google.com/drive/folders/1aDEr-jOR1JeqiV-N49DuED27sSKuCKwu?usp=drive_link

We test the model from checkpooint model_best_ep769_val_loss_0.0388.pth. The result are shown below:

	Eval	Test
Loss	0.0388	-
P5 acc	0.918	-
P10 acc	0.976	0.9948/0.9958
P15 acc	0.985	-
Dist	2.50	-

Instruction

Configuration

The configuration for training, evaluation and test is set on the configs/task.yaml, modify it according to your need.

Training

• You should change the data_dir setting in configs/taks.yaml to your dataset path, then execute the following command to start training:

python train.py --config_file task.yaml 

• This process will generate the metadata and cached_dataset directories which enable the fast data loading in training.
• The training script has already contain the testing phase using the model trained under all epoches.

Testing

• Running test.py scripy will get the tracking prediction on test split and store the result on submission.csv file

python test.py --config_file task.yaml --checkpoint your_model_path --log_dir store_result_path

Viualization

Using the following command to visualze the tracking result on eval/test split:

python visualize --config_file task.yaml --checkpoint your_model_path --split test

Other details

Acknowledgement

This code is based on the https://github.com/EETChallenge/challenge_demo_code