Home

Awesome

Learn-to-Decompose: Cascaded Decomposition Network for Cross-Domian Few-Shot Facial Expression Recignition

[Paper] [Code]

This is an official implementation of the following paper:

Learn-to-Decompose: Cascaded Decomposition Network for Cross-Domian Few-Shot Facial Expression Recignition

Xinyi Zou, Yan Yan*, Jing-Hao Xue, Si Chen, and Hanzi Wang

European Conference on Computer Vision (ECCV), 2022

In breif, we have the following contributions:

  1. propose CDNet which cascades several shared learn-to-decompose (LD) module via a sequential decomposition mechanism to obtain the general expression prototypes and their corresponing weights.
  2. develop a partial regularization strategy to combine the benefits of both episodic training and batch training
  3. achieve stat-of-the-art performance on various compound FER datasets under CD-FSL setting.

Please cite our paper if you find the code useful for your research.

@article{zou2022learn,
  title={Learn-to-Decompose: Cascaded Decomposition Network for Cross-Domain Few-Shot Facial Expression Recognition},
  author={Zou, Xinyi and Yan, Yan and Xue, Jing-Hao and Chen, Si and Wang, Hanzi},
  journal={arXiv preprint arXiv:2207.07973},
  year={2022}
}

Usage

Prerequisites

conda create --name py37 python=3.7
conda install pytorch torchvision -c pytorch
pip3 install -r requirements.txt

Install

Clone this repository:

git clone https://github.com/zouxinyi0625/CDNet.git
cd CDNet

Datasets

Training & Testing

We adopt the pretrained ResNet18 from here.

cd batch
python pretrain_e.py --dataset multi --name cascade_e --testset CFEE --split val --color 3 --w_domain 1.0
cd episodic
python train_l2d_pre.py --dataset multi --testset CFEE --split val --train_aug --weight net --name cascade_pre --w_d 1.0 --w_t 1.0 --pretrain

Note