STAR Loss: Reducing Semantic Ambiguity in Facial Landmark Detection.

Paper Link: arxiv | CVPR 2023

Pytorch implementation of Self-adapTive Ambiguity Reduction (STAR) loss.
STAR loss is a self-adaptive anisotropic direction loss, which can be used in heatmap regression-based methods for facial landmark detection.
Specifically, we find that semantic ambiguity results in the anisotropic predicted distribution, which inspires us to use predicted distribution to represent semantic ambiguity. So, we use PCA to indicate the character of the predicted distribution and indirectly formulate the direction and intensity of semantic ambiguity. Based on this, STAR loss adaptively suppresses the prediction error in the ambiguity direction to mitigate the impact of ambiguity annotation in training. More details can be found in our paper.

Dependencies

python==3.7.3
PyTorch=1.6.0
requirements.txt

Dataset Preparation

Step1: Download the raw images from COFW, 300W, and WFLW.
Step2: We follow the data preprocess in ADNet, and the metadata can be download from the corresponding repository.
Step3: Make them look like this:

# the dataset directory:
|-- ${image_dir}
   |-- WFLW
      | -- WFLW_images
   |-- 300W
      | -- afw
      | -- helen
      | -- ibug
      | -- lfpw
   |-- COFW
      | -- train
      | -- test
|-- ${annot_dir}
   |-- WFLW
      |-- train.tsv, test.tsv
   |-- 300W
      |-- train.tsv, test.tsv
   |--COFW
      |-- train.tsv, test.tsv

Usage

Work directory: set the ${ckpt_dir} in ./conf/alignment.py.
Pretrained model:

Dataset	Model
WFLW	google / baidu
300W	google / baidu
COFW	google / baidu

Training

python main.py --mode=train --device_ids=0,1,2,3 \
               --image_dir=${image_dir} --annot_dir=${annot_dir} \
               --data_definition={WFLW, 300W, COFW}

Testing

python main.py --mode=test --device_ids=0 \
               --image_dir=${image_dir} --annot_dir=${annot_dir} \
               --data_definition={WFLW, 300W, COFW} \
               --pretrained_weight=${model_path} \

Evaluation

python evaluate.py --device_ids=0 \
                   --model_path=${model_path} --metadata_path=${metadata_path} \
                   --image_dir=${image_dir} --data_definition={WFLW, 300W, COFW} \

Results

The models trained by STAR Loss achieved SOTA performance in all of COFW, 300W and WFLW datasets.

BibTeX Citation

Please consider citing our papers in your publications if the project helps your research. BibTeX reference is as follows.

@inproceedings{Zhou_2023_CVPR,
    author    = {Zhou, Zhenglin and Li, Huaxia and Liu, Hong and Wang, Nanyang and Yu, Gang and Ji, Rongrong},
    title     = {STAR Loss: Reducing Semantic Ambiguity in Facial Landmark Detection},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {15475-15484}
}

Acknowledgments

This repository is built on top of ADNet. Thanks for this strong baseline.

TimSC/STAR