Presentation 2022-01-24
Efficient dense interpolation of 4D light fields obtained by sparse viewpoints based on depth estimation
Hidemichi Yoshino, Kazuya Kodama, Takayuki Hamamoto,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) In order to easily enjoy immersive 3D visual environment reproducing light fields, inexpensive multi-view imaging systems are required.We have proposed a method for fast interpolation of sparse multi-view images by using parallel processing on GPGPUs, where dense light fields can be obtained without massive camera arrays. Actually, disparity estimation among multi-view images by plane-sweeping and simple epipolar plane image (EPI) synthesis allow us to reconstruct dense light fields in real time based on GPU-based view interpolation that flexibly works for various 3D scenes. In this paper, its implementation on GPGPUs is first described in detail, and then we discuss improvement of the previous method to achieve higher efficiency while avoiding degradation of reconstructed light fields. For EPI interpolation based on estimated disparity of sparse multi-view images, instead of the conventional analysis by simple pixel correspondence, we integrate pixel values on certain regions of the EPI by weighted summation using their consistency of disparity. Regardless of its computational cost, recent Tensor Cores enable us to expect significant speed-up of the novel EPI interpolation unlike exhaustive searches for pixel correspondence. Experimental results show that our novel view interpolation using weighted summation robustly obtains sufficient quality of reconstructed dense light fields.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) light field / interpolation / multi-view / parallel processing / GPGPU
Paper # IE2021-27
Date of Issue 2022-01-17 (IE)

Conference Information
Committee IE
Conference Date 2022/1/24(1days)
Place (in Japanese) (See Japanese page)
Place (in English) National Institute of Informatics
Topics (in Japanese) (See Japanese page)
Topics (in English) Image Processing, Image Coding, etc.
Chair Kazuya Kodama(NII)
Vice Chair Hiroyuki Bandoh(NTT) / Toshihiko Yamazaki(Univ. of Tokyo)
Secretary Hiroyuki Bandoh(KDDI Research) / Toshihiko Yamazaki(Nagoya Inst. of Tech.)
Assistant Shunsuke Iwamura(NHK) / Shinobu Kudo(NTT)

Paper Information
Registration To Technical Committee on Image Engineering
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Efficient dense interpolation of 4D light fields obtained by sparse viewpoints based on depth estimation
Sub Title (in English)
Keyword(1) light field
Keyword(2) interpolation
Keyword(3) multi-view
Keyword(4) parallel processing
Keyword(5) GPGPU
1st Author's Name Hidemichi Yoshino
1st Author's Affiliation Tokyo University of Science/National Institute of Informatics(Tokyo Univ. of Science/NII)
2nd Author's Name Kazuya Kodama
2nd Author's Affiliation National Institute of Informatics(NII)
3rd Author's Name Takayuki Hamamoto
3rd Author's Affiliation Tokyo University of Science(Tokyo Univ. of Science)
Date 2022-01-24
Paper # IE2021-27
Volume (vol) vol.121
Number (no) IE-346
Page pp.pp.1-4(IE),
#Pages 4
Date of Issue 2022-01-17 (IE)