OmniSCV: An Omnidirectional Synthetic Image Generator for Computer Vision.

Area of Science:

• Computer Vision
• Image Processing
• Robotics

Background:

• Omnidirectional and 360º images are increasingly prevalent, necessitating specialized computer vision algorithms due to inherent distortions.
• Training robust computer vision models requires large, annotated datasets, which are challenging to acquire for omnidirectional imagery.

Purpose of the Study:

• To present a tool for synthesizing omnidirectional image datasets with pixel-wise semantic and depth information.
• To address the need for high-quality, ground-truth data for training and testing omnidirectional computer vision algorithms.

Main Methods:

• Utilizing Unreal Engine 4 to generate realistic virtual environments and capture omnidirectional images.
• Implementing a diverse range of projection models, including central (equirectangular, cylindrical, fish-eye) and non-central systems.
• Providing pixel-accurate semantic, depth, and camera calibration data for synthesized images.

Main Results:

• The developed tool successfully generates photorealistic omnidirectional images with comprehensive ground-truth data.
• Demonstrated the tool's capability by validating various computer vision tasks, including line extraction, 3D layout recovery, SLAM, and 3D reconstruction.
• Achieved the first reported generation of photorealistic non-central omnidirectional images.

Conclusions:

• The proposed tool significantly facilitates the creation of datasets for omnidirectional computer vision research.
• Enables precise training and testing of algorithms for 3D vision and scene understanding using synthetically generated data.
• Offers a versatile solution for generating diverse omnidirectional image datasets, including novel non-central projection types.