Multimodal Cross-Domain Few-Shot Learning for Egocentric Action Recognition (ECCV'24)

[Paper][Supplementary][Project Page][Poster]

This is the official code release for our ECCV 2024 paper
"Multimodal Cross-Domain Few-Shot Learning for Egocentric Action Recognition".

🔨 Installation

# Create a virtual environment
python3 -m venv mm-cdfsl
source mm-cdfsl/bin/activate

# Install the dependencies
pip install -r requirements.txt

📂 Data Preparation

Training Split

You can find the train/val split files for all three target datasets in cdfsl folders.

Data Structure

Please follow the data structure as detailed in DATA_STRUCTURE.md.

📍 Model Zoo

You can brouse the checkpoints of pre-trained model, comparison methods, and our models in this folder or directly download from the following links.

Pre-Train

Method	Source Dataset	Target Dataset	Modality	Ckpt
VideoMAE	Kinetics-400	-	RGB	checkpoint
VideoMAE	Ego4D	-	RGB	checkpoint
VideoMAE w/ classifier	Ego4D	EPIC	RGB	checkpoint
VideoMAE w/ classifier	Ego4D	EPIC	flow	checkpoint
VideoMAE w/ classifier	Ego4D	EPIC	pose	checkpoint
VideoMAE w/ classifier	Ego4D	MECCANO	RGB	checkpoint
VideoMAE w/ classifier	Ego4D	MECCANO	flow	checkpoint
VideoMAE w/ classifier	Ego4D	MECCANO	pose	checkpoint
VideoMAE w/ classifier	Ego4D	WEAR	RGB	checkpoint
VideoMAE w/ classifier	Ego4D	WEAR	flow	checkpoint
VideoMAE w/ classifier	Ego4D	WEAR	pose	checkpoint

2nd Stage

Method	Source Dataset	Target Dataset	Modality	Ckpt
STARTUP++	Ego4D	EPIC	RGB	checkpoint
STARTUP++	Ego4D	MECCANO	RGB	checkpoint
STARTUP++	Ego4D	WEAR	RGB	checkpoint
Dynamic Distill++	Ego4D	EPIC	RGB	checkpoint
Dynamic Distill++	Ego4D	MECCANO	RGB	checkpoint
Dynamic Distill++	Ego4D	WEAR	RGB	checkpoint
CDFSL-V	Ego4D	EPIC	RGB	checkpoint
CDFSL-V	Ego4D	MECCANO	RGB	checkpoint
CDFSL-V	Ego4D	WEAR	RGB	checkpoint
Ours	Ego4D	EPIC	RGB, flow, pose	checkpoint
Ours	Ego4D	MECCANO	RGB, flow, pose	checkpoint
Ours	Ego4D	WEAR	RGB, flow, pose	checkpoint

🔥 Training

1. Pre-training

Please make sure that you set modality (e.g., rgb) and dataset (e.g., epic) in configs/trainer/pretrain_trainer.yaml and confings/data_module/pretrain_data_module.yaml.

python3 lit_main_pretrain.py train=True test=False

2. Multimodal Distillation

Please make sure that you set modality (e.g., rgb) and dataset (e.g., epic) in configs/trainer/mm_distill_trainer.yaml and confings/data_module/mm_distill_data_module.yaml. Also, you need to set the ckpt path of all modalities in configs/trainer/mm_distill_trainer.yaml.

python3 lit_main_mmdistill.py train=True test=False 

🔍 Evaluation

To evaluate the model in 5-way 5-shot setting with 600 runs, please run the following command.

python3 lit_main_mmdistill.py train=False test=False data_module.n_way=5 data_module.k_shot=5 data_module.episodes=600

✍️ Citation

If you use this code for your research, please cite our paper.

@inproceedings{Hatano2024MMCDFSL,
    author = {Masashi Hatano, Ryo Hachiuma, Ryo Fujii and Hideo Saito},
    title = {Multimodal Cross-Domain Few-Shot Learning for Egocentric Action Recognition},
    booktitle = {European Conference on Computer Vision (ECCV)},
    year = {2024},
}