Spatial-Temporal Graph Enhanced DETR Towards Multi-Frame 3D Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Detection Transformer (DETR) models have advanced CNN-based object detection.
• The application of DETR-like paradigms to multi-frame 3D object detection is underexplored.

Purpose of the Study:

• To present STEMD, an end-to-end framework enhancing DETR for multi-frame 3D object detection.
• To address challenges in spatial-temporal modeling and redundant predictions.

Main Methods:

• Introduced a spatial-temporal graph attention network for inter-object interactions and temporal dependencies.
• Incorporated previous frame outputs to initialize decoder queries, mitigating missing hard cases.
• Utilized an IoU regularization term to suppress similar, non-positive queries and reduce redundant boxes.

Main Results:

• Demonstrated effectiveness in challenging multi-frame 3D object detection scenarios.
• Achieved improved performance with only a minor increase in computational overhead.
• Successfully modeled complex object interactions and temporal dynamics.

Conclusions:

• STEMD offers a robust enhancement to DETR-like architectures for multi-frame 3D object detection.
• The proposed methods effectively handle spatial-temporal complexities and improve prediction accuracy.
• STEMD presents a computationally efficient solution for advanced 3D object detection tasks.