AD-DETR: DETR with asymmetrical relation and decoupled attention in crowded scenes

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Pedestrian detection in crowded scenes faces challenges with repeated predictions, high occlusion rates, and small visible pedestrian areas.
• Existing detection transformer (DETR) frameworks exhibit symmetry, leading to synchronous prediction updates and duplicate detections.

Purpose of the Study:

• To address the limitations of current pedestrian detection methods in crowded environments.
• To propose a novel detection framework, AD-DETR (asymmetrical relation detection transformer), to eliminate duplicate predictions and reduce false/missing detections.

Main Methods:

• Introduced an asymmetric relationship fusion mechanism to enable queries to learn to eliminate duplicate predictions by fusing relative relationships of surrounding predictions.
• Developed a decoupled cross-attention head to restrict attention range, focusing on visible regions and high-confidence areas, thus reducing noise from occluded objects.
• Integrated these modules into a DETR-based framework, utilizing deformable-DETR with ResNet-50 backbone.

Main Results:

• Achieved state-of-the-art performance on the CrowdHuman dataset with 92.6% average precision, 40.0% MR$^{-2}$, and 84.4% Jaccard index.
• Demonstrated significant improvements over existing methods like Iter-E2EDet and MIP.
• Showcased the robustness and effectiveness of the proposed method for crowded scene pedestrian detection.

Conclusions:

• The proposed AD-DETR framework effectively tackles key challenges in crowded scene pedestrian detection.
• The asymmetric relation fusion and decoupled cross-attention mechanisms significantly enhance detection accuracy and reduce errors.
• AD-DETR offers a robust and high-performing solution for query-based object detection in complex, crowded environments.