Stripe Sensitive Convolution for Omnidirectional Image Dehazing.

Area of Science:

• Computer Vision
• Image Processing
• Virtual Reality

Background:

• Haze degrades the quality of 360° photos/videos and immersive virtual reality (VR) experiences.
• Existing single image dehazing methods primarily focus on standard planar images, neglecting omnidirectional imagery.

Purpose of the Study:

• To develop a novel neural network pipeline for single omnidirectional image dehazing.
• To address distortion issues in equirectangular projections inherent to 360° imagery.

Main Methods:

• Creation of the first hazy omnidirectional image dataset, including synthetic and real-world samples.
• Introduction of a novel stripe sensitive convolution (SSConv) to calibrate distortion by extracting and weighting feature stripes.
• Design of an end-to-end network that jointly performs haze removal and depth estimation using SSConv, leveraging depth maps for global context.

Main Results:

• The proposed SSConv effectively handles distortion in omnidirectional images.
• The developed network achieves superior dehazing performance on synthetic and real-world datasets.
• Significant improvements in 3D object detection and 3D layout estimation for hazy omnidirectional images were demonstrated.

Conclusions:

• The novel pipeline and SSConv are effective for single omnidirectional image dehazing.
• The method enhances the quality of 360° imagery and improves downstream 3D perception tasks.
• This work advances the capabilities of processing and utilizing omnidirectional images in challenging atmospheric conditions.