CyberAgentAILab / SuperNormal

[CVPR 2024] Official implementation of "SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration"
MIT License
148 stars 7 forks source link

SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration

Xu Cao · Takafumi Taketomi
CyberAgent

CVPR 2024


arXiv PDF [//]: # ( ) [//]: # ( Project Page)

### Update - **2024/09/30**: Real-world raw data and step-by-step data pre-processing instructions are available. See [here](./data_capture_and_preprocessing/README.md).
Teaser Fast and fine-grained 3D reconstruction from multi-view surface normal maps.
### Quick Start Code was tested on Ubuntu 18.04 (WSL2) using Python 3.8, PyTorch 2.1.0, and CUDA 11.8 on an Nvidia RTX4090Ti (24GB). **Before started, please ensure CUDA is installed in your environment ([11.8 can be found here](https://developer.nvidia.com/cuda-11-8-0-download-archive?target_os=Linux&target_arch=x86_64&Distribution=WSL-Ubuntu&target_version=2.0&target_type=deb_local)).** It is required by [tiny-cuda-nn](https://github.com/NVlabs/tiny-cuda-nn).
You should see something like the following after typing `nvcc --version` ```commandline nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2022 NVIDIA Corporation Built on Wed_Sep_21_10:33:58_PDT_2022 Cuda compilation tools, release 11.8, V11.8.89 Build cuda_11.8.r11.8/compiler.31833905_0 ```
Clone the repository and prepare the conda environment: ```commandline git clone https://github.com/CyberAgentAILab/SuperNormal.git cd SuperNormal . ./create_env.sh ``` Download data (~1.8GB): ```commandline ./download_data.sh ``` Run on the DiLiGenT-MV benchmark objects or on our captured objects: ```commandline ./run_diligent.sh # Training should take about 50 seconds per object ``` or ```commandline ./run_own_object.sh # Training should take about 5 minutes per object ``` Results are saved under `./exp`. NOTE: If RuntimeError like below occurs, `apt install ninja-build` may resolve the error. ``` RuntimeError: Ninja is required to load C++ extensions ``` ### Hyperparameter tuning tips Training hyperparameters are defined in `./configs/*.conf`. Some important hyperparameters are: - `dataset.normal_dir`: You can choose normal maps estimated by different methods as input for DiLiGenT-MV benchmark objects. - `train.end_iter`: The number of iterations for training. Should be adjusted according to the number of views and normal map resolutions. - `train.increase_bindwidth_every`: A strategy used in [Neuralangelo](https://research.nvidia.com/labs/dir/neuralangelo/) to progressively activate finer hash grid during training. Less than `end_iter`/`model.encoding.n_levels` should be fine. - `train.batch_size`: Number of patches in each batch for training. Should be adjusted according to the GPU memory. - `train.patch_size`: Better to be fixed to 3, i.e., each patch is 3x3. Large patch size will cause inaccurate volume rendering results for boudary pixels in a patch. ### Modifications to NerfAcc We add several functions to the original [NerfAcc](https://www.nerfacc.com) to adapt it to patch-based volume rendering. The key new functions (which are indicated by `patch_based` in function name) are in [third_parties/nerfacc-0.3.5/nerfacc-0.3.5/nerfacc/cuda/csrc/render_weight.cu/](./third_parties/nerfacc-0.3.5/nerfacc-0.3.5/nerfacc/cuda/csrc/render_weight.cu) and [third_parties/nerfacc-0.3.5/nerfacc-0.3.5/nerfacc/vol_rendering.py](./third_parties/nerfacc-0.3.5/nerfacc-0.3.5/nerfacc/vol_rendering.py). ### Acknowledgement This repo is built up on [NeuS](https://github.com/Totoro97/NeuS) and benefits from the amazing [tiny-cuda-nn](https://github.com/NVlabs/tiny-cuda-nn) and [NerfAcc](https://www.nerfacc.com). We also learned a lot from [instant-nsr-pl](https://github.com/bennyguo/instant-nsr-pl). ### Citation If you find our work useful in your research, please consider citing: ```bibtex @inproceedings{supernormal2024cao, title={SuperNormal: {N}eural Surface Reconstruction via Multi-View Normal Integration}, author={Cao Xu and Taketomi Takafumi}, booktitle={CVPR}, year={2024} } ```