• 2024.7.25: Release inference code and checkpoint weight of Diffcalib in the repo.
• 2024.7.25: Release arXiv paper, with supplementary material.

conda create -n diffcalib python=3.10
conda activate diffcalib
pip install -r requirements.txt
pip install -e .

Download the intrinsic data from MonoCalib

First, Download the stable-diffusion-2-1 and transform it to the 8 inchannels from modify which will be put in --checkpoint, put the checkpoints under ./checkpoint/

Then, Download the pre-trained models diffcalib-best.zip from BaiduNetDisk(Extract code：xn5p). Please unzip the package and put the checkpoints under ./checkpoint/ which will be put in --unet_ckpt_path.

finally, you can run the bash to evaluate our model in the benchmark.

sh scripts/run_incidence_mutidata.sh

For depth and incident map visualization , download shift model from res101 (Extract code: g3yi), diffcalib-pcd.zip from BaiduNetDisk(Extract code:20z6) and install torchsparse packages as follows

sudo apt-get install libsparsehash-dev
pip install --upgrade git+https://github.com/mit-han-lab/[email protected]

Then we can reconstruct 3D shape from a single image.

bash scripts/run_incidence_depth_pcd.sh

• Marigold: Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation. arXiv, GitHub.

For non-commercial use, this code is released under the LICENSE.

@article{he2024diffcalib,
  title   = {DiffCalib: Reformulating Monocular Camera Calibration as Diffusion-Based Dense Incident Map Generation},
  author  = {Xiankang He and Guangkai Xu and Bo Zhang and Hao Chen and Ying Cui and Dongyan Guo},
  year    = {2024},
  journal = {arXiv preprint arXiv: 2405.15619}
}