Shadow-DETR: Alleviating matching conflicts through shadow queries.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• DETR (DEtection TRansformer) models achieve state-of-the-art performance using end-to-end object detection.
• One-to-one matching in DETR, while effective, suffers from slow convergence due to matching conflicts and limited supervision.

Purpose of the Study:

• To analyze and mitigate the limitations of one-to-one matching in DETR models.
• To enhance training efficiency and improve the overall performance of DETR-based object detection.

Main Methods:

• Introduced 'shadow samples' (negative samples resembling positive ones) and ignored their classification loss.
• Computed regression loss for shadow samples to provide additional localization supervision.
• Proposed a loss-balancing strategy and a feature-aware query initialization approach.

Main Results:

• The proposed Shadow-DETR strategy enhances network training efficiency and overall performance.
• Shadow-DETR substantially boosts existing methods like DAB-DETR, Deformable-DETR, and DINO.
• Achieved comparable performance with state-of-the-art (SOTA) methods.

Conclusions:

• Shadow-DETR effectively addresses conflicts and limited supervision issues in DETR's one-to-one matching.
• The proposed methods enhance DETR training dynamics and improve detection accuracy.
• Shadow-DETR offers a promising direction for advancing end-to-end object detection.