Authors: Vibhas K. Vats, Md. Alimoor Reza, David J. Crandall, and Soon-heung Jung

NOTE: Please contact Vibhas Vats (vkvatsdss@gamil.com) for any help.

Traditional multi-view stereo (MVS) methods primarily depend on photometric and geometric consistency constraints. In contrast, modern learning-based algorithms often rely on the plane sweep algorithm to infer 3D geometry, applying explicit geometric consistency (GC) checks only as a post-processing step, with no impact on the learning process itself. In this work, we introduce GC-MVSNet++, a novel approach that actively enforces geometric consistency of reference view depth maps across multiple source views (multi-view) and at various scales (multi-scale) during the learning phase (see Fig. 1 above). This integrated GC check significantly accelerates the learning process by directly penalizing geometrically inconsistent pixels, effectively halving the number of training iterations compared to other MVS methods. Furthermore, we introduce a densely connected cost regularization network with two distinct block designs—simple and feature-dense—optimized to harness dense feature connections for enhanced regularization. Extensive experiments demonstrate that our approach achieves a new state-of-the-art on the DTU and BlendedMVS datasets secure second place on the Tanks and Temples benchmark. To our knowledge, GC-MVSNet++ is the first method to enforce multi-view, multi-scale geometric consistency during learning.

Our code is tested with Python==3.9 and above, PyTorch==2.0 and above, CUDA==10.8 and above on LINUX systems with NVIDIA GPUs.

To use GC-MVSNet, clone this repo:

git clone 
cd GC-MVSNet

In GC-MVSNet, we mainly use DTU, BlendedMVS and Tanks and Temples to train and evaluate our models. You can prepare the corresponding data by following the instructions below.

For DTU training set, you can download the preprocessed DTU training data and Depths_raw (both from Original MVSNet), and unzip them to construct a dataset folder like:

dtu_training
 ├── Cameras
 ├── Depths
 ├── Depths_raw
 └── Rectified

For DTU testing set, you can download the preprocessed DTU testing data (from Original MVSNet) and unzip it as the test data folder, which should contain one cams folder, one images folder and one pair.txt file.

We use the low-res set of BlendedMVS dataset for both training and testing. You can download the low-res set from orignal BlendedMVS and unzip it to form the dataset folder like below:

BlendedMVS
 ├── 5a0271884e62597cdee0d0eb
 │     ├── blended_images
 │     ├── cams
 │     └── rendered_depth_maps
 ├── 59338e76772c3e6384afbb15
 ├── 59f363a8b45be22330016cad
 ├── ...
 ├── all_list.txt
 ├── training_list.txt
 └── validation_list.txt

Download our preprocessed Tanks and Temples dataset and unzip it to form the dataset folder like below:

tankandtemples
 ├── advanced
 │  ├── Auditorium
 │  ├── Ballroom
 │  ├── ...
 │  └── Temple
 └── intermediate
        ├── Family
        ├── Francis
        ├── ...
        └── Train

Set the configuration in train.sh:

• MVS_TRAINING: traing data path
• LOG_DIR: checkpoint saving path
• MASK_TYPE: mask name
• CONS2INCO_TYPE: type of averaging operatio for geometric mask
• AVERAGE_WEIGHT_GAP: decide the range of geometric penalty
• OPERATION: operation between penalty and per-pixel error
• nviews: nviews for the network, default 5
• GEO_MASK_SUM_TH: value of M
• R_DEPTH_MIN_THRESH: stage-wise RDD threshold
• DIST_THRESH: stage-wise PDE threshold
• LR: learning rate
• LREPOCHS: learning rate decay epochs:decay factor
• weight_decay: weight decay term
• EPOCHS: training epochs
• NGPUS: no of gpus
• BATCH_SIZE: batch size
• ndepths: stage-wise number of depth hypothesis planes
• DLOSSW: stage-wise weights for loss terms
• depth_interval_ratio: depth interval ratio

For a fair comparison with other SOTA methods on Tanks and Temples benchmark, we finetune our model on BlendedMVS dataset after training on DTU dataset.

Set the configuration in finetune_bld.sh:

Important Tips: to reproduce our reported results, you need to:

• compile and install the modified gipuma from Yao Yao as introduced below
• use the latest code as we have fixed tiny bugs and updated the fusion parameters
• make sure you install the right version of python and pytorch, use some old versions would throw warnings of the default action of align_corner in several functions, which would affect the final results
• be aware that we only test the code on 2080Ti and Ubuntu 18.04, other devices and systems might get slightly different results
• make sure that you use the gcmvsnet.ckpt for testing

NOTE: After all these, you still might not be able to exactly reproduce the results on DTU. The final results on DTU also depends on fusion hyperparameters and your hardware. But you should not get much different numbers from the report numbers.

To start testing, set the configuration in test_dtu.sh:

• TESTPATH: test directory
• TESTLIST: do not change
• FUSIBLE_PATH: path to fusible of fusibile
• CKPT_FILE: model checkpoint
• OUTDIR: output directory
• DEVICE_ID: GPU ID
• max_h=864 (Do not change)
• max_w=1152 (Do not change) Gipuma filter paramerters
• FILTER_METHOD: gipuma
• gipuma_prob_thresh: probability threshold for Fusibile
• gipuma_disparity_thresh: disparity threhold for Fusibile
• gipuma_num_consistenc: View consistency threshold for Fusibile
• depth_interval_ratio: Depth interval ratio
• ndepths: same as traning

Run:

bash test_dtu.sh

Note: we use the gipuma fusion method by default.

To install the gipuma, clone the modified version from Yao Yao. Modify the line-10 in CMakeLists.txt to suit your GPUs. Othervise you would meet warnings when compile it, which would lead to failure and get 0 points in fused point cloud. For example, if you use 2080Ti GPU, modify the line-10 to:

set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-O3 --use_fast_math --ptxas-options=-v -std=c++11 --compiler-options -Wall -gencode arch=compute_70,code=sm_70)

If you use other kind of GPUs, please modify the arch code to suit your device (arch=compute_XX,code=sm_XX). Then install it by cmake . and make, which will generate the executable file at FUSIBILE_EXE_PATH. Please note

For quantitative evaluation on DTU dataset, download SampleSet and Points. Unzip them and place Points folder in SampleSet/MVS Data/. The structure looks like:

SampleSet
├──MVS Data
      └──Points

In DTU-MATLAB/BaseEvalMain_web.m, set dataPath as path to SampleSet/MVS Data/, plyPath as directory that stores the reconstructed point clouds and resultsPath as directory to store the evaluation results. Then run DTU-MATLAB/BaseEvalMain_web.m in matlab.

We recommend using the finetuned models to test on Tanks and Temples benchmark.

Similarly, set the configuration in test_tnt.sh as described before. By default, we use dynamic filternation for TnT.

To generate point cloud results, just run:

bash test_tnt.sh

Note that：

• The parameters of point cloud fusion have not been studied thoroughly and the performance can be better if cherry-picking more appropriate thresholds for each of the scenes.
• The dynamic fusion code is borrowed from AA-RMVSNet.

For quantitative evaluation, you can upload your point clouds to Tanks and Temples benchmark.

@InProceedings{vats2023gcmvsnet,
    author    = {Vats, Vibhas K and Joshi, Sripad and Crandall, David and Reza, Md. and Jung, Soon-heung },
    title     = {GC-MVSNet: Multi-View, Multi-Scale, Geometrically-Consistent Multi-View Stereo},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2024},
    pages     = {3242--3252},
}
@InProceedings{vats2024gcmvsnet++,
    author    = {Vats, Vibhas K and Crandall, David and Reza, Md. and Jung, Soon-heung },
    title     = {Blending 3D Geometry and Machine Learning for Multi-View Stereopsis},
    booktitle = {xxx},
    month     = {January},
    year      = {2024},
    pages     = {0000},
}

We borrow some code from CasMVSNet, TransMVSNet. We thank the authors for releasing the source code.