NeRF-Editing: Geometry Editing of Neural Radiance Fields

Abstract

Implicit neural rendering, especially Neural Radiance Field (NeRF), has shown great potential in novel view synthesis of a scene. However, current NeRF-based methods cannot enable users to perform user-controlled shape deformation in the scene. While existing works have proposed some approaches to modify the radiance field according to the user's constraints, the modification is limited to color editing or object translation and rotation. In this paper, we propose a method that allows users to perform controllable shape deformation on the implicit representation of the scene, and synthesizes the novel view images of the edited scene without re-training the network. Specifically, we establish a correspondence between the extracted explicit mesh representation and the implicit neural representation of the target scene. Users can first utilize well-developed mesh-based deformation methods to deform the mesh representation of the scene. Our method then utilizes user edits from the mesh representation to bend the camera rays by introducing a tetrahedra mesh as a proxy, obtaining the rendering results of the edited scene. Extensive experiments demonstrate that our framework can achieve ideal editing results not only on synthetic data, but also on real scenes captured by users.

Environment

Install jittor
Other dependencies (click to expand)
- opencv_python==4.5.2.52
- imageio==2.17.0
- trimesh==3.9.8
- numpy==1.19.2
- pyhocon==0.3.57
- icecream==2.1.0
- tqdm==4.50.2
- scipy==1.7.0
- PyMCubes==0.1.2
- natsort==8.1.0
- tensorboardX-2.5

We also use the pyrender to get the depth map.

pip install pyrender

Download OpenVolumeMesh to the OpenVolumeMesh folder
Download Eigen to the volumeARAP_batch/Eigen folder

Data preparation

Suppose the image data is in the $data_dir/images folder, we first estimate the camera poses with colmap. Then we process the camera poses with the command

python process_colmap.py $data_dir $data_dir

Finally the data folder looks like

$data_dir
├── colmap_output.txt (colmap output)
├── database.db (colmap output)
├── images ($data_dir/images)
├── intrinsics.txt
├── pose
├── rgb
└── sparse (colmap output)

We now provide a provide a ready-to-use dataset hbychair collected by ourselves in google drive, along with the pre-trained model and deformation results. You can put the data into the data folder.

Or you can use nerf-synthetic dataset directly, see ./confs/wmask_lego.conf as an example.

Technological process

Training

we adopt the training strategy of NeuS.

python exp_runner.py --mode train --conf ./confs/womask_hbychair.conf --case hbychair_neus

Extract mesh

python exp_runner.py --mode validate_mesh --conf ./confs/womask_hbychair.conf --case hbychair_neus --is_continue # use latest checkpoint

We have provided a simplified mesh mesh_nofloor_simp.obj

Render image before editing

python exp_runner.py --mode circle --conf ./confs/womask_hbychair_render.conf --case hbychair_neus --is_continue  --obj_path ./logs/hbychair_wo_mask/mesh_nofloor_simp.obj

Note: obj_path is optional, which provides better rendering results.

Construct cage mesh

python exp_runner.py --mode validate_mesh --conf ./confs/womask_hbychair.conf --case hbychair_neus --is_continue --do_dilation

We have provided a cage mesh mesh_cage_nofloor.obj

Construct tetrahedral mesh using TetWild.

./TetWild ../../src/logs/hbychair_wo_mask/mesh_cage_nofloor.obj

Note that we modify the tetrahedra storage format of Tetwild output. Therefore, please compile the tetwild in this repository following the instructions here.

Change the output to `ovm` format.

./simple_mesh ../../src/logs/hbychair_wo_mask/mesh_cage_nofloor_.txt ../../src/logs/hbychair_wo_mask/mesh_cage_nofloor_.ovm

simple_mesh can be obtained using the CMakeLists.txt in the OpenVolumeMesh folder.

Editing

Deform the extracted mesh with any mesh editing tool, and put the (sequence) mesh in $deformed_dir folder.

We have provided a deformed mesh deformed_mesh.obj and a folder named as mesh_seq

Propagate editing

Generate the controlpoint.txt to guide the deformation.

python barycentric_control_pts_jittor.py

Note that specify the mesh_path (extracted mesh), tet_path (tetrahedra mesh) and deformed_dir(deformed mesh sequence) first.

And the format of controlpoint.txt is listed below.

10 (Number of sequence)
N (Num of control points)
x1 y1 z1
x2 y2 z2
...
N (Num of control points)
x1 y1 z1
x2 y2 z2
...
.
.
.
N (Num of barycentric coordinate)
id1 id2 id3 id4 (vert index of this tet)
u1 v1 w1 z1
id1' id2' id3' id4'
u2 v2 w2 z2
...

Compile the volumeARAP_batch project to obtain volumeARAP, and deform the tetrehedra mesh.

./volumeARAP ../../src/logs/hbychair_wo_mask/mesh_cage_nofloor_.ovm ../../src/logs/hbychair_wo_mask/mesh_seq/2_barycentric_control.txt ../../src/logs/hbychair_wo_mask/mesh_seq_ovm 0

Rendering after editing

python exp_runner.py --mode circle --conf ./confs/womask_hbychair_render.conf --case hbychair_neus --is_continue --use_deform --reconstructed_mesh_file ./logs/hbychair_wo_mask/mesh_cage_nofloor_.txt --deformed_mesh_file ./logs/hbychair_wo_mask/mesh_seq_ovm/arap_result_0000_.ovm --obj_path ./logs/hbychair_wo_mask/deformed_mesh.obj

fix camera (generate sequential editing results in a fixed camera)

python exp_runner.py --mode circle --conf ./confs/womask_hbychair_render.conf --case hbychair_neus --is_continue --use_deform --reconstructed_mesh_file ./logs/hbychair_wo_mask/mesh_cage_nofloor_.txt --deformed_mesh_file ./logs/hbychair_wo_mask/mesh_seq_ovm/arap_result_0000_.ovm --obj_path ./logs/hbychair_wo_mask/deformed_mesh.obj --fix_camera

Acknowledgement

This code borrows heavily from NeuS.

Citation

If you found this code useful please cite our work as:

@inproceedings{yuan2022nerf,
  title={NeRF-editing: geometry editing of neural radiance fields},
  author={Yuan, Yu-Jie and Sun, Yang-Tian and Lai, Yu-Kun and Ma, Yuewen and Jia, Rongfei and Gao, Lin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={18353--18364},
  year={2022}
}

suzhoushr/NeRF-Editing