Fast Rolling Shutter Correction in the Wild.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Rolling shutter (RS) distortion is a common artifact in uncalibrated videos.
• Existing methods often require explicit camera motion and depth estimation for correction.
• These intermediate steps can be computationally intensive and limit real-time applications.

Purpose of the Study:

• To develop an efficient and effective rolling shutter correction (RSC) method for uncalibrated videos.
• To enable real-time distortion correction without explicit camera parameter estimation.
• To improve the quality of corrected videos for downstream 3D analysis tasks.

Main Methods:

• A novel point-wise rolling shutter correction (RSC) technique is proposed.
• This method implicitly rectifies distorted pixels by rescaling optical flow.
• A direct RS correction (DRSC) framework handles pixel-wise varying distortions.

Main Results:

• The proposed approach achieves real-time performance (40fps for 480p) on CPU.
• Demonstrated superiority over state-of-the-art methods in effectiveness and efficiency across diverse scenarios.
• Corrected videos showed improved suitability for 3D analysis tasks like visual odometry and structure-from-motion.

Conclusions:

• The developed DRSC framework offers an efficient, uncalibrated solution for rolling shutter correction.
• The method effectively handles complex distortions from camera motion, object movement, and scene depth.
• The output quality enhances subsequent 3D computer vision applications.