This is the official implementation for the 2023 Transactions on Multimedia (Early Access) paper, MosaicMVS: Mosaic-based Omnidirectional Multi-view Stereo for Indoor Scenes.
As we cannot share the entire dataset, only provide some pictures of scene1 for testing.
For testing our code, run the shell file:
We run this code using GeForce RTX 3090 in following environment.
python >=3.7
Pytorch ==1.10
CUDA >=9.0
This dataset consists of 11 scenes.
Parts of the scene 1 RGB images in SAOI dataset are uploaded in scene1 folder for implementation.
For more datasets you needed, please contact us.
If you capture customed mosaic images, you must run the COLMAP to obtain undistorted image, and camera parameters. You need following datas with your images,
Undistorted images (RGB)
Intrinsic
Extrinsics
View selection text (pair.txt)
To run this code,
git clone https://github.com/min-jung-shin/MosaicMVS.git
sh test_mosaic.sh
To fusion output depthmaps,
CUDA_VISIBLE_DEVICES=0,1 python fusioncas.py -data <dir_of_depths> --pair <dir_of_pair> --vthresh 3 --pthresh .8,.8,.8 --outmask <dir_of_masks> --out_dir <dir_of_output pointcloud>
For reconstrction evaluation, refer to the code in the python fusioncas.py.
("total.mean: ", sum(total)/len(total))
To evaluate depth map, COLMAP sparse reconstruction depthmaps are needed. You can run the COLMAP.
Thanks to Xiaodong Gu for opening source of his excellent work CasMVSNet. Thanks to Jingyang Zhang of his excellent work Vis-MVSNet.