Probabilistic Intersection-Over-Union for Training and Evaluation of Oriented Object Detectors.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Oriented object detection is a challenging problem, with existing methods often adapting Horizontal Bounding Box (HBB) detectors.
• Current Oriented Bounding Box (OBB) approaches face difficulties with complex formulations, custom backpropagation, and ambiguous representations for irregular or circular objects.

Purpose of the Study:

• To develop a novel, unified approach for training, representing, and evaluating oriented object detectors.
• To address the limitations of traditional OBBs, particularly for ambiguous object shapes.

Main Methods:

• Introduced Gaussian Bounding Boxes (GBBs) as fuzzy representations for oriented objects.
• Proposed a Hellinger distance-based similarity metric between GBBs, yielding a differentiable closed-form expression for a localization loss.
• Demonstrated that GBBs naturally map to Elliptical Bounding Boxes (EBBs), resolving ambiguity for circular objects.

Main Results:

• The proposed metric, termed Probabilistic Intersection-over-Union (ProbIoU), correlates strongly with the Intersection-over-Union (IoU) of corresponding EBBs.
• Using ProbIoU as a regression loss achieves competitive results against state-of-the-art methods without extra hyperparameters or custom implementations.
• ProbIoU shows promise as an evaluation metric for oriented object detectors.

Conclusions:

• Gaussian Bounding Boxes (GBBs) provide a robust and flexible representation for oriented objects.
• The Probabilistic Intersection-over-Union (ProbIoU) metric offers an effective and simplified solution for training and evaluating oriented object detectors.