DPE-MVS

This is python wrapper of DPE-MVS from [https://github.com/ckh0715/DPE-MVS].

Setup

as a project with uv

This guide is aimed for linux users.

1. Install DPE-MVS dependencies.

sudo apt install -y libopencv-dev

2. Install uv

curl -LsSf https://astral.sh/uv/install.sh | sh

3. Build venv with uv

uv sync

as a python library in uv projects.

1. Install DPE-MVS dependencies.

sudo apt install -y libopencv-dev

2. Add this project as a dependency.

Edit root pyproject.toml as following

[project]
dependencies = [
   "DPE_MVS"
]
[tool.uv.sources]
DPE_MVS = { git = "<url/of/this/git-reop>" }

as a C++ build by cmake

sudo apt install -y cmake libopencv-dev
mkdir build
cd build
cmake ../csrc/DPE-MVS
make
cd ..

How to call functions.

Prepare DPE

python src/DPE_MVS/colmap2mvsnet.py --dense_folder <path/to/colmap/model/folder> --save_folder <path/to/DPEprepared/folder>

Run DPE as a python library

from DPE_MVS import dpe_mvs

dpe_mvs(
   dense_folder="path/to/DPEprepared/folder",  # type: str. 
   gpu_index=0,        # Optional. type: int. default = 0.
   verbose=True    # Optional. If True, log will be printed.
   fusion=False,   # Optional. type: bool, default = False. If True, point clouds file (.ply) will be created, which causes overhead.
   viz=False,   # Optional. type: bool, default = False. If True, all visualization info will be saved, which causes overhead.
   depth=True,   # Optional. type: bool, default = False. If True, depth.npy will be created, which causes overhead.
   normal=False,   # Optional. type: bool, default = False. If True, normal.npy will be created, which causes overhead.
   weak=False,   # Optional. type: bool, default = False. If True, weak.npy will be created, which causes overhead.
   edge=False   # Optional. type: bool, default = False. If True, edge.npy will be created, which causes overhead.
)

Run DPE with C++ build

Args and their order are the same as using as python library (above).

./build/DPE "path/to/DPEprepared/folder" 0 true false false true false false false

Outputs (other than visualization)

depth.npy

Depth map.
shape: [height, width]
type: np.float32

normal.npy

Normal map.
shape: [height, width, 3]
type: np.float32

weak.npy

Hold the confidence of each pixel. 0 is no confidence (should be discarded), 1 is weak confidence, 2 is strong confidence.
shape: [height, width]
type: np.int8

edge.npy

Hold edges. 0 is not edge, 1 is on the edge.
shape: [height, width]
type: np.int8