GAOC: A Gaussian Adaptive Ochiai Loss for Bounding Box Regression.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Bounding box regression (BBR) loss is crucial for object detection accuracy.
• Current BBR loss functions based on Intersection over Union (IoU) have limitations in handling predicted box scale and the drift problem.
• Existing methods lack scale invariance and effective positional deviation handling.

Purpose of the Study:

• To introduce a novel BBR loss function, Gaussian Adaptive Ochiai BBR loss (GAOC), to overcome limitations of existing methods.
• To enhance object detection robustness and accuracy by addressing scale effects and positional deviations.
• To provide a more effective solution for bounding box regression in computer vision.

Main Methods:

• Developed GAOC by combining the Ochiai Coefficient (OC) for scale invariance and a Gaussian Adaptive (GA) distribution for positional similarity.
• The OC component normalizes bounding box dimensions, ensuring scale invariance.
• The GA distribution models coordinate distances to reduce sensitivity to positional deviations.

Main Results:

• GAOC was integrated into YOLOv5 and RT-DETR object detection models.
• Evaluated on PASCAL VOC and MS COCO 2017 benchmarks.
• GAOC consistently outperformed existing BBR loss functions in experimental evaluations.

Conclusions:

• The proposed GAOC loss function offers superior performance compared to existing BBR loss functions.
• GAOC effectively addresses scale invariance and positional drift issues in bounding box regression.
• This novel approach enhances overall object detection accuracy and robustness.