Multi-parallax views synthesis for three-dimensional light-field display using unsupervised CNN.

Area of Science:

• Computer Vision
• Computer Graphics
• Display Technology

Background:

• Multi-view applications, particularly Three-Dimensional (3D) displays, are widely used.
• Capturing dense multiple views for 3D light-field displays remains challenging.
• View synthesis offers a viable alternative for generating necessary views.

Purpose of the Study:

• To propose an unsupervised network, Multi-Parallax View Net (MPVN), for synthesizing multi-parallax views for 3D light-field display.
• To address the difficulties in obtaining training targets and synthesizing views at arbitrary positions.
• To enable high-quality prediction of virtual views at any parallax position between existing views.

Main Methods:

• Developed an unsupervised network (MPVN) for synthesizing multi-parallax views.
• Re-projected existing parallax views to target positions to create input towers.
• Network processes input towers, outputting color and selection towers for final image reconstruction via per-pixel weight summing.
• Employed end-to-end unsupervised training to minimize prediction errors.

Main Results:

• MPVN successfully synthesizes multi-parallax views for 3D light-field display.
• Achieved high-quality virtual view prediction at arbitrary parallax positions.
• Structural Similarity Index (SSIM) for synthetic views consistently exceeded 0.95 in experiments.

Conclusions:

• The proposed MPVN method effectively synthesizes multi-parallax views for 3D light-field displays.
• This approach overcomes limitations of traditional view synthesis methods.
• MPVN shows promise for providing sufficient views for future 3D light-field display applications.