Bézier Palm: A Free Lunch for Palmprint Recognition

1. Tencent Youtu Lab, 2. UCLA, 3. Hefei University of Technology, 4. Shanghai Jiaotong University.

Abstract:

Palmprints are private and stable information for biometric recognition. In the deep learning era, the development of palmprint recognition is limited by the lack of sufficient training data. In this paper, by observing that palmar creases are the key information to deep-learning-based palmprint recognition, we propose to synthesize training data by manipulating palmar creases. Concretely, we introduce an intuitive geometric model which represents palmar creases with parameterized Bézier curves. By randomly sampling Bézier parameters, we can synthesize massive training samples of diverse identities, which enables us to pretrain large-scale palmprint recognition models. Experimental results demonstrate that such synthetically pretrained models have a very strong generalization ability: they can be efficiently transferred to real datasets, leading to significant performance improvements on palmprint recognition. For example, under the open-set protocol, our method improves the strong ArcFace baseline by more than 10% in terms of [email protected] And under the closed-set protocol, our method reduces the equal error rate (EER) by an order of magnitude.

Method:

Observation and motivation:

Creases synthesis with Bézier curves:

What's new

Synthesized Samples:

Synthesize in 2D:

The figure below provides some synthesized samples, each row contains sample of the same identity.

Synthesize in 3D:

TBD

Experimental Results

[email protected] curve on public datasets.

[email protected] curve of different methods. AF, MF, R50 denotes ArcFace, MobileFaceNet and ResNet-50. Hover (desktop) or click (mobile) to see numbers.

Our method consistently outperforms the baseline with substantial margin.

ImageNet pretrained v.s. our synthetically pretrained.

[email protected] curve of models pretrained with ImageNet dataset and our synthesized samples. Hover (desktop) or click (mobile) to see numbers.

Compared to ImageNet pretraining, our synthetically pretrained models generalize better when finetuning on real palmprint recognition datasets.

FAQ:

If you have any questions about our paper, please do not hesitate to leave comments below.

Citation:

If our methods are helpful to your research, please kindly consider to cite:
@article{zhao2022geometric,
  title={Geometric Synthesis: A Free lunch for Large-scale Palmprint Recognition Model Pretraining},
  author={Zhao, Kai and Shen, Lei and Zhang, Yingyi and Zhou, Chuhan and Wang, Tao and Zhang, Ruixin and Ding, Shouhong and Jia, Wei and Shen, Wei},
  journal={arXiv preprint arXiv:2203.05703},
  year={2022}
}
@article{shen2022distribution,
  title={Distribution Alignment for Cross-device Palmprint Recognition},
  author={Shen, Lei and Zhang, Yingyi and Zhao, Kai and Zhang, Ruixin and Shen, Wei},
  url={https://data.kaizhao.net/publications/wildpalm2022.pdf},
  year={2022}
}