REDNet

implementation of REDNet (CVPR 2020)

“A Novel Recurrent Encoder-Decoder Structure for Large-Scale Multi-view Stereo Reconstruction from An Open Aerial Dataset”(arxiv：https://arxiv.org/abs/2003.00637)

The proposed network was trained and tested on a single NVIDIA TITAN RTX 2080Ti (24G).

Requirements

CUDA 10.0
cudnn 7.6
python 3.6
tensorflow-gpu 1.13.1
numpy 1.18
opencv-python 4.1

Data Preparation

Download the WHU MVS dataset. http://gpcv.whu.edu.cn/data/WHU_dataset/WHU_MVS_dataset.zip.
Unzip the dataset to the WHU_MVS_dataset folder.

Train

In “train.py”, set data_root to your train data path YOUR_PATH/WHU_MVS_dataset/train
Train REDNet (TITAN RTX 2080Ti 24G):

python train.py

Test

Download the pre-trained REDNet model (https://pan.baidu.com/s/13BfLJ3sNfQL_933wZjR8PA, code：ohqx)
Unzip it to MODEL_FOLDERfolder.
In test.py, set dense_folder to your test data path YOUR_PATH/WHU_MVS_dataset/test, set model_dir to your model path MODEL_FOLDER, set depth sample number max_d.
Run REDNet：

Python test.py

The test outputs were stored in YOUR_PATH/WHU_MVS_dataset/test/depths_rednet/, including depth map XXX_init.pfm, probability map XXX_prob.pfm, scaled images XXX.jpg and camera parameters XXX.txt.

Predict

If you want to apply REDNet to your own data, please structure your data into such a folder.

Image Files

All image files are stored in the Images folder.

Camera files

All camera files are stored in the Camsfolder.
The text file contains the camera extrinsic, intrinsic and the depth range:

extrinsic
E00 E01 E02 E03
E10 E11 E12 E13
E20 E21 E22 E23
E30 E31 E32 E33

f(pixel)  x0(pixel)  y0(pixel)

DEPTH_MIN   DEPTH_MAX   DEPTH_INTERVAL
IMAGE_INDEX 0 0 0 0 WIDTH HEIGHT

Make sure the camera extrinsic is Twc [Rwc|twc], camera orientation is XrightYup.

In viewselect.py, set dense_folder to your data path.
Run:

Python viewselect.py

The output file viewpair.txt will stored in your data path.

In predict.py, set dense_folder to your data path. set model_dir to your model path MODEL_FOLDER, set depth sample number max_d and image size max_w, max_h.
Run:

Python predict.py

The outputs were stored in YOUR_DATA_PATH/depths_rednet/.

We provided the script fusion.py to apply depth map filter for post-processing, which converted the per-view depth maps to 3D point cloud. 5. In fusion.py, set dense_folder to your data path. 6. Run:

Python fusion.py

Final point clouds are stored in YOUR_DATA_PATH/rednet_fusion/.

Reference

This project is based on the implementation of R-MVSNet. Thanks Yao, Yao for providing the source code (https://github.com/YoYo000/MVSNet)

yGaoJiany/REDNet