This is a method proposed for 2022 Ego4D Looking-at-Me Challenge, utilizing Ego4D dataset and an optical flow dataset, this dataset is created with the RGB images in Ego4D by a neural network called FlowFormer.

conda create -n egoflow python=3.8 
conda activate egoflow
conda install pytorch=1.7.1 torchvision=0.8.2 cudatoolkit=11.0
pip install -r requirements.txt

1. FlowFormer

Create a folder

mkdir checkpoints

and put sintel.pth in it.

2. A pre-trained EgoFlow model for testing

1. Download Ego4D dataset following the official guidance.

2. Run all "*_odd.py" and "*_even.py" files to generate the optical flows. Please check the roots of datasets in these python files.

python run.py --model GazeLSTM --exp_path output_train --num_workers 16 --batch_size 64 --gamma 0.5

python run.py --eval --checkpoint checkpoint_EgoFlow.pth --model GazeLSTM --exp_path output_test --num_workers 16 --batch_size 128