ESSENCE-Net

TPAMI-2025 Revisiting Supervised Learning-based Photometric Stereo Networks

Xiaoyao Wei, Zongrui Li, Binjie Ding, Boxin Shi, Xudong Jiang, Gang Pan, Yanlong Cao, and Qian Zheng https://ieeexplore.ieee.org/document/10948383

Dependencies

ESSENCE-Net is implemented in PyTorch with Ubuntu 18.04 and an NVIDIA GeForce RTX 3090 GPU (24GB).

• Python 3.8.5
• PyTorch (version = 1.90)
• numpy
• scipy
• CUDA-11.1
• einops

Testing

Test on the DiLiGenT dataset

• Dense setups (96 input images)

python main_test.py --in_img_num 96

• Sparse setups (10 input images)

python main_test.py --in_img_num 10

Test on the DiLiGenT10² dataset, DiLiGenT-Π dataset, and DiLiGenRT dataset

The ground truth of the DiLiGenT10² dataset, DiLiGenT-Π dataset, and DiLiGenRT dataset is not publicly available. You can use these codes to estimate normal maps and submit the estimated normal maps to the corresponding website for evaluation of normal errors.

Training

The training code will be made available soon.

Results on the DiLiGenT dataset, DiLiGenT10² dataset, DiLiGenT-Π dataset, and DiLiGenRT dataset

We have provided the estimated surface normal maps (error maps) on the DiLiGenT, DiLiGenT10², DiLiGenT-Π, and DiLiGenRT benchmark datasets under 96/100 input images in ./pre_trained_model/.

Acknowledgement

Our code is partially based on https://github.com/guanyingc/PS-FCN.