A Review of DEtection TRansformer: From Basic Architecture to Advanced Developments and Visual Perception...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Detection Transformer (DETR) pioneered end-to-end object detection using Transformers, removing manual components like anchor boxes and Non-Maximum Suppression (NMS).
• Original DETR faced challenges including slow convergence, inadequate small object detection, and low computational efficiency, necessitating further research and development.

Purpose of the Study:

• To systematically review the technical evolution of DETR models from a problem-driven perspective.
• To analyze advancements in core components such as attention mechanisms, query design, training strategies, and architectural efficiency.
• To provide a comprehensive overview of DETR's applications and future research directions.

Main Methods:

• A systematic literature review focusing on problem-driven advancements in DETR.
• Analysis of key technical improvements including attention mechanisms, query formulation, and training methodologies.
• Categorization of DETR applications across various domains like autonomous driving and medical imaging.

Main Results:

• Significant progress has been made in improving DETR's convergence speed, small object detection accuracy, and overall efficiency.
• Key advancements have been observed in attention mechanisms, query optimization, and training strategies.
• DETR has demonstrated broad applicability in diverse fields, including autonomous driving, medical imaging, and remote sensing.

Conclusions:

• The DETR architecture has evolved substantially, addressing its initial limitations through targeted innovations.
• Continued research is exploring extensions to fine-grained classification and video understanding.
• This review provides a structured understanding of DETR's development, highlighting current challenges and future research avenues.