End-to-End Object Detection with Transformers

DETR reframed object detection as direct set prediction: learned query slots propose boxes in parallel, Hungarian matching assigns one positive owner to each target, and unmatched slots learn no object. Follow the mechanism, the decoder ablations, and the scale-dependent COCO evidence.

Reading focus: How a fixed bank of learned object queries represents a variable-sized set through the no-object class. Why one-to-one Hungarian matching turns duplicate predictions into training errors instead of an inference-time cleanup problem. What the 42.0 AP headline hides about large and small objects, long training, and fixed query capacity.

ECCV 2020. Nicolas Carion et al.. 13 min read, hard difficulty.