appearance-scanner
About
This repository is an implementation of the neural network proposed in Free-form Scanning of Non-planar Appearance with Neural Trace Photography
For any questions, please email xiaohema98 at gmail.com
Usage
System Requirement
- Windows or Linux(The codes are validated on Win10, Ubuntu 18.04 and Ubuntu 16.04)
- Python >= 3.6.0
- Pytorch >= 1.6.0
- tensorflow>=1.11.0, meshlab and matlab are needed if you process the test data we provide
Training
- move to appearance_scanner
- run train.bat or train.sh according to your own platform
Notice that the data generation step
python data_utils/origin_parameter_generator_n2d.py %data_root% %Sample_num% %train_ratio%
should be run only once.
Training Visulization
When training is started, you can open tensorboard to observe the training process. There will be two log images of a certain training sample, one is the sampled lumitexels from 64 views and the other is an composite image from six images in the order of groundtruth lumitexel, groundtruth diffuse lumitexel, groundtruth specular lumitexel, predicted lumitexel, predicted diffuse lumitexel and predicted specular lumitexel.
Trained lighting pattern will also be showed. Trained model will be found in the log_dir
set in train.bat/train.sh.
License
Our source code is released under the GPL-3.0 license for acadmic purposes. The only requirement for using the code in your research is to cite our paper:
@article{Ma:2021:Scanner,
author = {Ma, Xiaohe and Kang, Kaizhang and Zhu, Ruisheng and Wu, Hongzhi and Zhou, Kun},
title = {Free-Form Scanning of Non-Planar Appearance with Neural Trace Photography},
year = {2021},
issue_date = {August 2021},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
volume = {40},
number = {4},
issn = {0730-0301},
url = {https://doi.org/10.1145/3450626.3459679},
doi = {10.1145/3450626.3459679},
journal = {ACM Trans. Graph.},
month = jul,
articleno = {124},
numpages = {13},
keywords = {illumination multiplexing, SVBRDF, optimal lighting pattern}
}
For commercial licensing options, please email hwu at acm.org. See COPYING for the open source license.
Reconstruction process
The reconstruction needs photographs taken with our scanner, a pre-trained network model and a pre-captured geometry shape as input. First, perform structure-from-motion with COLMAP, resulting in a 3D point cloud and camera poses with respect to it. Next, this point cloud is precisely aligned with the pre-captured shape. Then the view information of each vertex can be assembled as the input of the network. Last, we fit the predicted grayscale specular lumitexel with L-BFGS-B, to obtain the refletance parameters.
Download our Cheongsam test data and unzip it in appearance_scanner/data/.
Three sample photographs captured from the Cheongsam object. The brightness of the original images has been doubled for a better visualization.
Download our model and unzip it in appearance_scanner/.
1. Camera Registration
SFM/run.bat
first to brighten the raw images
1.1 Run 1.2 Open Colmap and do the following steps
1.2.1 New project
1.2.2 Feature extraction
Copy the parameters of our camera in device_configuration/cam.txt to Custom parameters
.
1.2.3 Feature matching
Tick guided_matching
and run.
1.2.4 Reconstruction options
Do not tick multiple_models
in the General sheet.
Do not tick refine_focal_length/refine_extra_params/use_pba
in the Bundle sheet.
Start reconstruction.
1.2.5 Bundle adjustment
Do not tick refine_focal_length/refine_principal_point/refine_extra_params
.
1.2.6
Make a folder named undistort_feature in Cheongsam/ and export model as text in undistort_feature folder. Three files including cameras.txt, images.txt and point3D.txt will be saved.
1.2.7
Dense reconstruction -> select undistort_feature folder -> Undistortion -> Stereo
Since we upload all the photos we taken, it will take a long time to run this step. We strongly recommend you run
colmap stereo_fusion --workspace_path path --input_type photometric --output_path path/fused.ply
//change path to undistort_feature folder
when the files amount in undistort_feature/stereo/normal_maps arise to around 200-250. It will output a coarse point cloud in undistort_feature/ .
Delete the noise points and the table plane.
Save fused.ply.
2. Extract measurements
move your own model to models/
and run appearance_scanner/test_files/prepare_pattern.bat
run extract_measurements/run.bat
3. Align mesh
3.1 Use meshlab to align mesh roughly
Open fused.ply and Cheongsam/scan/Cheongsam.ply in the same meshlab window. Cheongsam.ply is pre-capptured with a commercial mobile 3D scanner, EinScan Pro 2X Plus.
Align two mesh and save project file in Cheongsam/scan/Cheongsam.aln, which records the transform matrix between two meshes.
run CoherentPointDrift/run.bat
to align Cheongsam.ply to fused.ply.
3.2 Further Alignment
run CoherentPointDrift/CoherentPointDrift-master/simplify/run.bat
to simplify two meshes. It will call meshlabserver to simplify two meshes so that save the processing time.
Open the CPD project in Matlab and run main.m
.
After alignment done, run CoherentPointDrift/run_pass2.bat
. meshed-poisson_obj.ply will be saved in undistort_feature/ .
You should open fused.ply and meshed-poisson_obj.ply in the same meshlab window to check the quality of alignment. It is a key factor in the final result.
4. Generate view information from registrated cameras
4.1 Remesh
run ACVD/aarun.bat
save undistort_feature/meshed-poisson_obj_remeshed.ply as undistort_feature/meshed-poisson_obj_remeshed.obj
It is not necessary to reconstruct all the vertices on the pre-captured shape in our case. The remesh step will output an optimized 3D triangular mesh with a user defined vertex budget, which is controlled by NVERTICES
in aarun.bat
.
4.2 uvatlas
copy data_processing/device_configuration/extrinsic.bin to undistort_feature/ copy Cheongsam/512.exr and 1024.exr to undistort_feature/
run generate_texture/trans.bat
to transform mesh from colmap frame to world frame in our system and generate uv maps.
We recommend that you generate uv maps with resolution of 512x512 because it will save a lot of time and retain most details. The resolution of the results in our paper is 1024x1024.
You can set UVMAP_WIDTH
and UVMAP_HEIGHT
to 1024 in uv/uv_generator.bat
if you pursue higher quality.
4.3 Compute view information
Downloads embree and copy bin/embree3.dll, glfw3.dll, tbb12.dll to generate_texture/.
Downloads opencv and copy opencv_world#v.dll to generate_texture/. We use opencv3.4.3 in our project.
in generate_texture/texgen.bat
, set TEXTURE_RESOLUTION
to the certain resolution
choose the same line or the other reference on meshed-poisson_obj_remeshed.obj and on the physical object, then meature the lengths of both. Set the results to COLMAP_L
and REAL_L
. REAL_L
in mm.
The marker cylinder's diameter is 10cm, so we set REAL_L
to 100.
run generate_texture/texgen.bat
to output view information of all registrated cameras.
5. Gather data
run gather_data/run.bat
to gather the inputs to the network for each valid pixel on the texture map. A folder named images_{resolution} will be made in Cheongsam/.
6. Fitting
- Change %DATA_ROOT% and %TEXTURE_MAP_SIZE% in
fitting/tf_ggx_render/run.bat
. Then runfitting/tf_ggx_render/run.bat
. - A folder named
fitting_folder_for_server
will be generated under texture_{resolution}. - Upload the entire folder generated in previous step to a linux server.
- Change current path of terminal to
fitting_folder_for_server\fitting_temp\tf_ggx_render
, then runsplit.sh
orsplit1024.sh
according to the resolution you chosen. (split.sh is for 512. If you want to use custom texture map resolution, you may need to modify the $TEX_RESOLUTION in split.sh) - When the fitting procedure finished, a folder named
Cheongsam/images_{resolution}/data_for_server/data/images/data_for_server/fitted_grey
will be generated. It contains the final texture maps, including normal_fitted_global.exr, tangent_fitted_global.exr, axay_fitted.exr, pd_fitted.exr and ps_fitted.exr.
Note: If you find the split.sh cannot run properly and complain about abscent which_server argument, it's probably caused by the difference of linux and windows. Reading in the sh file and writing it with no changing of content on sever can fix this issue.
diffuse | specular | roughness |
normal | tangent |
7. Render results
We use the anisotropic GGX model to represent reflectance. The object can be rendered with path tracing using NVIDIA OptiX or openGL.
Reference & Third party tools
Shining3D. 2021. EinScan Pro 2X Plus Handheld Industrial Scanner. Retrieved January, 2021 from https://www.einscan.com/handheld-3d-scanner/2x-plus/
Colmap: https://demuc.de/colmap/
Coherent Point Drift: https://ieeexplore.ieee.org/document/5432191
ACVD: https://github.com/valette/ACVD
Embree: https://www.embree.org/
OpenCV: https://opencv.org/