A Systematic IoU-Related Method: Beyond Simplified Regression for Better Localization.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Current object detection models often employ simplified localization losses, such as Smooth-L1 Loss.
• These losses are inconsistent with the standard Intersection over Union (IoU) evaluation metric.
• Directly using IoU as a loss is problematic due to issues like zero gradients for non-overlapping boxes.

Purpose of the Study:

• To develop a novel loss function for object detection that aligns with the IoU evaluation metric.
• To improve bounding box localization accuracy without significantly increasing computational overhead.
• To address the limitations of existing localization losses and direct IoU application.

Main Methods:

• Proposed an Extended Intersection over Union (EIoU) metric, which is well-defined for non-overlapping boxes.
• Introduced a Convexification Technique (CT) to create a trainable loss function based on EIoU.
• Developed a Steady Optimization Technique (SOT) for smoother convergence.
• Integrated an IoU-predicting head to further enhance localization performance.

Main Results:

• The proposed method, integrated with Faster R-CNN (ResNet50+FPN), achieved significant performance gains.
• Demonstrated improvements of 4.2 mAP on VOC2007 and 2.3 mAP on COCO2017 over Smooth-L1 Loss.
• Showcased substantial gains at stricter metrics (AP90), with 8.2 mAP on VOC2007 and 5.4 mAP on COCO2017.
• Achieved these improvements with negligible increases in training and inference computational costs.

Conclusions:

• The proposed EIoU-based loss function effectively addresses the limitations of traditional localization losses.
• The novel approach significantly enhances bounding box regression accuracy in object detection.
• The method offers a computationally efficient way to boost performance, particularly under stricter evaluation criteria.