Focal DETR: Target-Aware Token Design for Transformer-Based Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Transformer-based object detection models face challenges with target attribute diffusion.
• Existing methods struggle to precisely localize and represent target features.

Purpose of the Study:

• To propose a novel target-aware token design for transformer-based object detection.
• To enhance feature representation and reduce interference from non-target regions.

Main Methods:

• Introduced a target-aware sampling module with four specialized patterns (small, large, vertical, horizontal).
• Developed a target-aware key-value matrix for direct feature map weighting.
• Proposed Focal DETR, a new variant of transformer-based object detection.

Main Results:

• Focal DETR achieved 44.7 AP on the COCO 2017 test set.
• Demonstrated superior performance compared to DETR and deformable DETR.
• Achieved 2.7 AP and 0.9 AP higher than DETR and deformable DETR, respectively.

Conclusions:

• The proposed target-aware token design effectively addresses target attribute diffusion.
• Focal DETR offers a significant improvement in object detection accuracy.
• The method shows strong potential for advancing transformer-based object detection.