/unsupervised_inter_intra_landmark

Primary LanguagePython

Unsupervised Landmark Learning with Inter-Intra Subject Consistencies

[paper]

Weijian Li, Haofu Liao, Shun Miao, Le Lu, and Jiebo Luo

IAPR International Conference on Pattern Recognition (ICPR), 2020, Oral Presentation

Introduction

We propose an unsupervised learning approach to image landmark discovery by incorporating the inter-subject landmark consistencies on facial images.

The proposed core model block can be found here

Prerequisites

  • Python 3.6
  • Pytorch 1.4

Preparation

  • CelebA dataset: please download the CelebA dataset, unzip and place it under ./celeba. Please also copy the file list_landmarks_align_celeba.txt to this repo's path ./.

  • AFLW dataset: please download the AFLW dataset, unzip and place it under ./aflw_release-2.

  • The pretrained checkpoint by ESanchezLozano is placed in ./checkpoint_fansoft/fan_109.pth which is a landmark detector pretrained on MPII human joint detection.

Training

To train the model, first train on the CelebA dataset:

python train.py --data_path celeba/Img/img_align_celeba_hq/ --cuda 1 --bSize 32 --num_workers 4

Testing

The trained model is saved at ./Exp_xxx. To test the trained model, first we need to extract the detected results on target datasets, i.e. AFLW or MAFL, for both training and test partitions. The default number of keypoints N=10:

python extract_data.py -f Exp_354 -e 33 -c checkpoint_fansoft/fan_109.pth -d MAFL-train --data_path celeba/Img/img_align_celeba_hq/ --cuda 1
python extract_data.py -f Exp_354 -e 33 -c checkpoint_fansoft/fan_109.pth -d MAFL-test --data_path celeba/Img/img_align_celeba_hq/ --cuda 1

Then we can train a linear regressor and compute NME errors:

python -f Exp_354 -e 33 -d MAFL -r 0.0001

Reference

If you find our paper and repo useful, please cite our paper. Thanks!

@article{li2020unsupervised,
  title={Unsupervised Learning of Landmarks based on Inter-Intra Subject Consistencies},
  author={Li, Weijian and Liao, Haofu and Miao, Shun and Lu, Le and Luo, Jiebo},
  journal={arXiv preprint arXiv:2004.07936},
  year={2020}
}

Related Project

Structured Landmark Detection via Topology-Adapting Deep Graph Learning

Weijian Li, Yuhang Lu, Kang Zheng, Haofu Liao, Chihung Lin, Jiebo Luo, Chi-Tung Cheng, Jing Xiao, Le Lu, Chang-Fu Kuo, Shun Miao

European Conference on Computer Vision (ECCV), 2020, [paper]

Credits

The code is based on previous work: Object landmark discovery through unsupervised adaptation. Thanks a lot for the sharing!