A Review of DEtection TRansformer: From Basic Architecture to Advanced Developments and Visual Perception Applications
Liang Yu, Lin Tang, Lifeng Mu
Abstract
DEtection TRansformer (DETR) introduced an end-to-end object detection paradigm using Transformers, eliminating hand-crafted components like anchor boxes and Non-Maximum Suppression (NMS) via set prediction and bipartite matching. Despite its potential, the original DETR suffered from slow convergence, poor small object detection, and low efficiency, prompting extensive research. This paper systematically reviews DETR's technical evolution from a "problem-driven" perspective, focusing on advancements in attention mechanisms, query design, training strategies, and architectural efficiency. We also outline DETR's applications in autonomous driving, medical imaging, and remote sensing, and its expansion to fine-grained classification and video understanding. Finally, we summarize current challenges and future directions. This "problem-driven" analysis offers researchers a comprehensive and insightful overview, aiming to fill gaps in the existing literature on DETR's evolution and logic.