A Maximum a Posteriori Estimation Framework for Robust High Dynamic Range Video Synthesis.

Area of Science:

• Computer Vision
• Image Processing
• Video Synthesis

Background:

• High dynamic range (HDR) image synthesis is an active research area.
• Extending HDR synthesis to video presents challenges in accurate object correspondence estimation due to data loss from poor exposures and varying intensity.
• Conventional optical flow methods are often inaccurate for HDR video due to these challenges.

Purpose of the Study:

• To develop a robust method for high dynamic range (HDR) video synthesis.
• To address the limitations of conventional optical flow methods in estimating object correspondence in HDR video.
• To achieve high-quality HDR video output with an improved complexity-performance tradeoff.

Main Methods:

• A statistical approach using maximum a posteriori estimation is proposed, avoiding exact correspondence estimation.
• The problem is reduced to an optimization problem for foreground and background estimation.
• Background is obtained via rank minimization, and foreground is estimated using a novel multiscale adaptive kernel regression technique.

Main Results:

• The proposed algorithm demonstrates superior performance in delivering high-quality HDR videos compared to state-of-the-art methods.
• Experimental results on real and synthetic datasets validate the algorithm's effectiveness.
• A complexity analysis shows a better complexity-performance tradeoff than conventional methods.

Conclusions:

• The novel statistical approach effectively synthesizes high-quality HDR videos.
• The method overcomes the limitations of traditional optical flow techniques for HDR video.
• The algorithm offers an improved balance between computational complexity and performance.