Dynamic Label Assignment for Object Detection by Combining Predicted IoUs and Anchor IoUs.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Label assignment is crucial for object detection model performance.
• Anchor-based models rely on Intersection over Union (IoU) thresholds to distinguish positive and negative samples.
• Existing methods use fixed or distribution-based adaptive thresholds.

Purpose of the Study:

• To introduce a dynamic label assignment approach using model predictions.
• To improve the selection of high-quality positive samples in object detection.
• To reduce the discrepancy between classification and IoU scores.

Main Methods:

• Implementing a dynamic label assignment strategy based on training status and predictions.
• Utilizing model predictions to select samples with higher IoU to ground truth boxes.
• Evaluating the impact of the adaptive label assignment on detection model performance.

Main Results:

• The proposed approach dynamically assigns labels based on predictions.
• More high-quality samples with higher IoUs are selected as positive samples.
• Improvements in detection model performance and reduced bounding box losses were observed.

Conclusions:

• Dynamic label assignment using predictions enhances object detection.
• The method effectively selects higher-quality positive samples, leading to better boundary box predictions.
• This adaptive strategy offers a simple yet effective improvement for object detection models.