About

This is a modified version of COLMAP 3.5 to take in multi-channel input data.

It is for the paper Learning Efficient Photometric Feature Transform for Multi-view Stereo

---

Requirements

The combination below is demonstrated. Other versions may cause compiling problems.

libs	version
cuda	10.0.130 / 11.0.2
cudnn	7.6.3.30 / 8.0.5.39
gcc/g++	5
cgal	4.7
boost	1.58
ceres	1.14.0

TIPS:

• For Ubuntu 20+, no entry of gcc/g++ lower than 7 is provided. One should add xenial repository as described here.
• Source codes of demonstrated version of ceres and CGAL are provided in the folder named 3rd_packages. If you use these codes, the recommended order of installing is: gcc/g++ -> cuda/cudnn -> boost -> ceres -> cgal
• If your gpu is 3090, add this flag to cmake to specify GPU architecture: -DCUDA_ARCHS=Ampere

---

Installation

Ubuntu

1. Install required libs by:

   sudo apt-get install \
       git \
       cmake \
       build-essential \
       libboost-program-options-dev \
       libboost-filesystem-dev \
       libboost-graph-dev \
       libboost-regex-dev \
       libboost-system-dev \
       libboost-test-dev \
       libeigen3-dev \
       libsuitesparse-dev \
       libfreeimage-dev \
       libgoogle-glog-dev \
       libgflags-dev \
       libglew-dev \
       qtbase5-dev \
       libqt5opengl5-dev \
       libcgal-dev
   
   sudo apt-get install libcgal-qt5-dev
   sudo apt-get install libatlas-base-dev libsuitesparse-dev
   

   Note: By automatic installation, the version of boost may not meet the version required.
2. Install Ceres Solver
   First unpack the source code of Ceres in 3rd_packages, then:

   cd ceres-solver
   git checkout $(git describe --tags) # Checkout the latest release
   mkdir build
   cd build
   cmake .. -DBUILD_TESTING=OFF -DBUILD_EXAMPLES=OFF
   make -j
   sudo make install
   

3. Compile COLMAP

   mkdir build
   cd build
   CC=/usr/bin/gcc-5 CXX=/usr/bin/g++-5 cmake ..
   make -j16
   sudo make install
   

   Note: If you have Anaconda installed, the compiler may complain the missing library of libtiff.so. That is because Anaconda changes the system PATH and hides the required libtiff.so.5. You need to find this library file and manually link that file to /usr/lib/x86_64-linux-gnu/libtiff.so.5

---

Usage

Since we only modified the channel of images for patch match, only the Undistort Image & Patch Match Stereo parts are different. The other parts are same with original COLMAP.

Channel of input images

The channel of input images is specified in line 50 of src/mvs/patch_match_cuda.cu. If you want to use a different channel number, recompilation is required.

Undistort

```
colmap image_undistorter --image_path $IMAGE_PATH --input_path $INPUT_PATH 
```

$IMAGE_PATH contains the multi-channel images. The name of multi-channel images should be same as the images used in SFM and an extra .bin is appended. For example, if the image names used in SFM are 0.png, 1.png,... 10.png. Then there should be 0.png.bin, 1.png.bin,... 10.png.bin in $IMAGE_PATH. The format of each bin file is:
3 integers: image_width, image_height, image_channel
(iamge_width * image_height * image_channel) floats (row major)

$INPUT_PATH contains the dense folder exported after SFM, by using the order in COLMAP GUI: file -> export as txt. The folder should contains 3 txt files: cameras.txt, images.txt and points.txt

Patch Match Stereo

```
colmap image_undistorter --workspace_Path $INPUT_PATH --PatchMatchStereo.multi_channel 1
```

$INPUT_PATH is same as above.