The official implementation of the paper Learning Efficient Photometric Feature Transform for Multi-view Stereo.
torch_renderer
Please clone this repo. As specified in tame.py(line 14), the repo should be in the same folder of training codes. The related calibration file can be found here.
training data
Download the pregenerated training data from here. Extract the data to the folder specified in train_horizontal.sh. In tame.py, the training program will visualize feature space every 5000 steps, the related data(required by tame.py in line 270) can be found here.
multi-channel COLMAP
Please refer to this repo to install the modified COLMAP for testing.
- run train_horizontal.sh to train Intensity-sensitive Branch
- run train_vertical.sh to train Intensity-insensitive Branch
- run train_combine.sh to load the pretrained model and jointly train the network
Please download the data and trained model:
Use val_files/inference_dift/infer_dift_codes.bat to generate feature maps. A folder named feature_maps will be generated in the folder of each objects. We provide pre-generated COLMAP dense folder for each objects, named as undistort_feature_dift. Please undistort the multi-channel featuremaps and run dense reconstruction with modifed multi-channel COLMAP.
colmap image_undistorter --image_path /path/to/feature/map/folder --input_path /path/to/undistort_feature_dif/ --output_path /path/to/undistort_feature_dif/ --input_type BIN
colmap patch_match_stereo --workspace_path . --PatchMatchStereo.multi_channel 1 --PatchMatchStereo.geom_consistency 1 --PatchMatchStereo.sigma_spatial 15 --PatchMatchStereo.sigma_color 5.0 --PatchMatchStereo.num_samples 20 --PatchMatchStereo.ncc_sigma 1.0
Each object has trans_mat_2_gt.txt containing a rigid transform matrix to transform reconstruction to groudtruth frame. You can use val_files/fusion_transfer_2_gt_cmp.py to produce point cloud and calculate reconstruction quality. Note that the results may slightly differ from our reported values in the paper, because of randomality of PCA, COLMAP reconstruction and alignment.
There is a roi.txt in each object's folder, one can specify this when undistorting images to accelerate dense matching.
Please checkout the brach diligent_mv and then follow the instruction in README.md.
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:
@InProceedings{Kang_2021_ICCV,
author = {Kang, Kaizhang and Xie, Cihui and Zhu, Ruisheng and Ma, Xiaohe and Tan, Ping and Wu, Hongzhi and Zhou, Kun},
title = {Learning Efficient Photometric Feature Transform for Multi-View Stereo},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2021},
pages = {5956-5965}
}
For commercial licensing options, please email hwu at acm.org.
See COPYING for the open source license.
Note: The reconstruction evaluation tool, ETH3DMultiViewEvaluation.exe and corresponding DLLs in val_files are not bounded with this license.