Dual-GDNet: Dual Guided-diffusion Network for Stereo Image Dense Matching

Neural Network Project for Ruei-Ping's Master Thesis

Author: Ruei-Ping Wang (Jacky)
Organizer: Yi-Chen Chen (Jack) and Ruei-Ping Wang (Jacky)

Size of parameter: 11,593,136

Prerequisite

• Windows 10 x64
• RTX 2080 Ti
• Python 3.6
• PyTorch v1.6
• CUDA Toolkit v10.2
• cuDNN v8
• VS2019 (msvc)

Dependency Compilation in Windows (gdnet-lib)

cd ./GANet/extensions/
rm -r build

replace CppExtension with CUDAExtension in line 5 of setup.py
python ./setup.py install

Installation

.env/Scripts/activate.bat
python -m pip install --upgrade pip
pip install wheel
pip install -r requirements.txt
pip3 install torch==1.10.1+cu102 torchvision==0.11.2+cu102 torchaudio===0.10.1+cu102 -f https://download.pytorch.org/whl/cu102/torch_stable.html

Issues

1. pytorch/pytorch#42467
2. pytorch/pytorch#11004 (comment)
3. https://blog.csdn.net/tanmx219/article/details/100829920

Modify pytorch package

修改 C:\Python37\Lib\site-packages\torch\include\c10\macros\Macros.h CONSTEXPR_EXCEPT_WIN_CUDA -> CONSTEXPR_EXCEPT_WIN_CUDA const

修改 C:\Python37\Lib\site-packages\torch\include\pybind11\cast.h 1449行 explicit operator type&() { return *(this->value); -> explicit operator type&() { *((type *)(this->value)); }

setup.py 使用CUDAExtension

Original KITTI 2015's height and width

height: 1242 = 2*3^3*23
width: 375 = 3*5^3

Efficiency

Dual-GANet: 13 seconds per KITTI 2015 image with size 384, 1248

Size of input image for models

The size of input images are required to be multiple for specific model.

Name	Height	Width	Max disparity
GDNet_sdc6	mod 64	mod 64	mod 32
GDNet_sdc6f	mod 64	mod 64	mod 32
GDNet_mdc6	mod 32	mod 32	mod 16
GDNet_mdc6f	mod 32	mod 32	mod 16
GDNet_fdc6	mod 32	mod 32	mod 8
GDNet_fdc6f	mod 32	mod 32	mod 8
LEAStereo_fdcf	mod 24	mod 24	mod 8
GDNet_sd9d6	mod 64	mod 64	mod 32

Evaluation

Name	Height	Width	Max disparity
Flyingthings3D	384	960	144
KITTI 2015	352	1216	144

Other

• Disparity with float32 must multiply 256 to be uint16 format