CCDet: Confidence-Consistent Learning for Dense Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detectors use classification scores to estimate localization quality.
• An inconsistency exists between classification scores and localization accuracy, leading to unreliable predictions.
• This inconsistency stems from inaccurate Intersection over Union (IoU) estimation and spatial misalignment of features.

Purpose of the Study:

• To propose a Confidence-Consistent Detector (CCDet) that addresses the confidence inconsistency in object detection.
• To improve the reliability of detection results for downstream applications.
• To enhance the performance of object detection models.

Main Methods:

• Developed Distribution-based IoU Prediction (DIP) for stable and accurate IoU estimation by learning the IoU probability distribution.
• Introduced Consistency-aware label assignment (CLA) using prediction performance and sample consistency for positive sample selection.
• Guided classification and localization tasks to promote similar feature distributions.

Main Results:

• CCDet effectively mitigates the confidence inconsistency between classification and localization.
• Achieved stable performance improvements across various object detection baselines.
• Attained a 50.1% AP on the MS COCO test-dev set with a single model and single scale, outperforming existing detectors.

Conclusions:

• The proposed CCDet successfully resolves the confidence inconsistency issue in object detection.
• CCDet offers a robust approach for enhancing the accuracy and reliability of object detection systems.
• The method demonstrates significant advancements in object detection performance, particularly on benchmark datasets.