NAN-DETR: noising multi-anchor makes DETR better for object detection

Area of Science:

• Computer Vision
• Robotics

Background:

• Object detection is vital for robotic vision, but current methods struggle with one-to-many matching.
• Transformer-based models require improvements for enhanced accuracy and robustness.

Purpose of the Study:

• To introduce NAN-DETR (Noising Multi-Anchor Detection Transformer), an advanced framework for object detection.
• To improve the one-to-many matching strategy in transformer-based object detection.

Main Methods:

• Proposed NAN-DETR, a framework built upon DETR (Detection Transformer).
• Implemented a decoder-based multi-anchor strategy for improved object localization.
• Introduced a centralization noising mechanism to mitigate anchor conflicts and enhance robustness.
• Integrated Complete Intersection over Union (CIoU) loss for more precise bounding box predictions.

Main Results:

• The multi-anchor strategy significantly enhanced detection accuracy.
• The noising mechanism increased model robustness by reducing anchor conflicts.
• CIoU loss led to more precise bounding box predictions compared to standard IoU loss.

Conclusions:

• NAN-DETR provides a highly reliable solution for object detection tasks.
• While not optimized for real-time performance, the framework demonstrates exceptional detection accuracy and robustness.