Bilateral Reference for High-Resolution Dichotomous Image Segmentation

Peng Zheng^1,4,5,6, Dehong Gao², Deng-Ping Fan^1*, Li Liu³, Jorma Laaksonen⁴, Wanli Ouyang⁵, Nicu Sebe⁶

¹Nankai University ²Northwestern Polytechnical University ³National University of Defense Technology ⁴Aalto University ⁵Shanghai AI Laboratory ⁶University of Trento

This repo holds the official weights of BiRefNet_lite trained in 2K resolution.

Training Sets:

DIS5K (except DIS-VD)
HRS10K
UHRSD
P3M-10k (except TE-P3M-500-NP)
TR-humans
AM-2k
AIM-500
Human-2k (synthesized with BG-20k)
Distinctions-646 (synthesized with BG-20k)
HIM2K
PPM-100

HR samples selection:

size_h, size_w = 1440, 2560
ratio = 0.8
h, w = image.shape[:2]
h >= size_h and w >= size_w or (h > size_h * ratio and w > size_w * ratio)

Validation Sets:

DIS-VD
TE-P3M-500-NP

Performance:

Dataset	Method	maxFm	wFmeasure	MAE	Smeasure	meanEm	HCE	maxEm	meanFm	adpEm	adpFm	mBA	maxBIoU	meanBIoU
DIS-VD	BiRefNet_lite-2K-general--epoch_232	.867	.831	.045	.879	.919	952	.925	.858	.916	.847	.796	.750	.739
TE-P3M-500-NP	BiRefNet_lite-2K-general--epoch_232	.993	.986	.009	.975	.986	.000	.993	.985	.833	.873	.825	.921	.891

Check the main BiRefNet model repo for more info and how to use it:
https://huggingface.co/ZhengPeng7/BiRefNet/blob/main/README.md

Remember to set the resolution of input images to 2K (2560, 1440) for better results when using this model.

Also check the GitHub repo of BiRefNet for all things you may want:
https://github.com/ZhengPeng7/BiRefNet

Acknowledgement:

Many thanks to @freepik for their generous support on GPU resources for training this model!

Citation

@article{zheng2024birefnet,
  title={Bilateral Reference for High-Resolution Dichotomous Image Segmentation},
  author={Zheng, Peng and Gao, Dehong and Fan, Deng-Ping and Liu, Li and Laaksonen, Jorma and Ouyang, Wanli and Sebe, Nicu},
  journal={CAAI Artificial Intelligence Research},
  volume = {3},
  pages = {9150038},
  year={2024}
}