| Title |
Handcrafted Cost based Multiview Stereo Matching Network |
| Authors |
전윤배(Yoonbae Jeon) ; 정태현(Taehyeon Jeong) ; 박인규(In Kyu Park) |
| DOI |
https://doi.org/10.5573/ieie.2022.59.8.61 |
| Keywords |
Multi-view stereo; CNN; Encoder-decoder network; Parallel processing; OpenCL |
| Abstract |
Multi-view stereo matching algorithm is a method of obtaining precise depth information by utilizing camera parameters from multiple input images taken from an arbitrary viewpoint. Since previous multi-view stereo matching algorithms are computationally complex, most of the implementation environment were limited to high-performance hardware. Additionally, previous multi-view stereo matching algorithms work well only when there is the little occluded area between adjacent images, and they are vulnerable to noise in the image. In this paper, we propose a method for estimating a dense depth map from a small number of input images through a CNN-based encoder-decoder network. The proposed method consists of the cost volume construction algorithm, the view-selection method for the target image, and the disparity map estimation method through a cost volume regression network. In this thesis, we show the possibility of implementing multi-view stereo matching with an OpenCL-based parallel algorithm in a low-power embedded board through the proposed algorithm and recover the precise depth map from a small number of images through a network trained with multiple image dataset. |