IoU Regression with H+L-Sampling for Accurate Detection Confidence.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object detection frameworks typically assume consistent data distributions for classification and bounding box regression.
• Existing sampling strategies for localization quality estimation are either inconsistent or complex to train.

Purpose of the Study:

• To propose an effective and simple sampling strategy for training the quality estimation branch in object detection.
• To improve the alignment between classification confidence and localization accuracy.

Main Methods:

• Introduced H+L-Sampling strategy, selecting high and low Intersection over Union (IoU) samples simultaneously.
• Developed accurate detection confidence by combining classification probability and localization accuracy for Non-Maximum Suppression (NMS) ranking.

Main Results:

• The H+L-Sampling strategy effectively trains the quality estimation branch, inheriting benefits of consistent sampling while reducing training complexity.
• Demonstrated improved detection performance through extensive experiments, validating the method's effectiveness.

Conclusions:

• The proposed H+L-Sampling strategy offers a more efficient and effective approach to training object detection models.
• Accurate detection confidence enhances the ranking mechanism in NMS, leading to superior overall detection accuracy.