Real-Time Light Field Depth Estimation via GPU-Accelerated Muti-View Semi-Global Matching

Yuanqi Wang, Xuesong Zhang, Hengsong Li, Anlong Ming

Abstract

The structured and redundant imagery of light field cameras can provide more robust depth estimation results while on the other hand demands a huge computation power, which limits its real-time applications, such as online industrial monitoring, 3D endoscopic surgery etc. This paper extends the classical SGM(Semi-global matching) [1] algorithm to light filed multi-view stereo framework, which can acquire sub-pixel level disparity estimation to cope with the micro-baseline of light field cameras. The whole algorithm is tailored for parallelization on GPU exploiting multi-stream asynchronization, multi-thread allocation, and multi-type memory management. The results show that our method’s execution time is less than 50ms on NVIDIA 1080Ti, which to our knowledge is the fastest among reported geometry based methods while keeping a comparative accuracy performance.
BibTeX
@inproceedings{wang2019real,
  author = {Yuanqi Wang and Xuesong Zhang and Hengsong Li and Anlong Ming},
  title = {{Real-Time Light Field Depth Estimation via GPU-Accelerated Muti-View Semi-Global Matching}},
  booktitle = {ICIP},
  year = {2019},
  doi = {10.1109/icip.2019.8803400},
}

Loading related content...

Figure