Field-of-View IoU for Object Detection in 360° Images.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• 360° cameras are increasingly popular, generating images in equirectangular projection (ERP).
• Standard object detection models struggle with ERP images due to inherent distortions.
• Existing methods require significant adaptation for effective 360° object detection.

Purpose of the Study:

• To propose novel techniques for enhancing object detection in 360° images.
• To address performance degradation caused by ERP image distortions.
• To provide methods easily integrable with existing object detection frameworks.

Main Methods:

• Field-of-View Intersection over Union (FoV-IoU): A metric for bounding box comparison in spherical images, suitable for training, inference, and evaluation.
• 360Augmentation: A data augmentation strategy involving random rotation of spherical images to mitigate sphere-to-plane projection biases.

Main Results:

• The proposed FoV-IoU and 360Augmentation techniques significantly improve object detection performance in 360° images.
• Consistent effectiveness demonstrated across various perspective object detectors.
• Experiments conducted on a 360° indoor dataset validated the methods' efficacy.

Conclusions:

• The introduced FoV-IoU and 360Augmentation are effective solutions for 360° object detection.
• These methods offer a practical approach to improve accuracy without redesigning existing detectors.
• The techniques provide a robust framework for advancing computer vision tasks in immersive imaging.