DETR (DEtection TRansformer), released by a Facebook AI Research team led by Nicolas Carion in May 2020, reframed object detection as a direct set prediction problem. Instead of the multi-stage pipelines that dominated detection, it used a transformer encoder-decoder to reason globally about the image and emit the final set of objects in one shot.
The conceptual win was removing hand-designed machinery. Classic detectors relied on anchor boxes (a grid of candidate shapes tuned per dataset) and non-maximum suppression, a post-processing step that deletes duplicate detections of the same object. DETR eliminated both. It predicts a fixed-size set of boxes and uses a set-based global loss with bipartite matching, which forces each ground-truth object to be matched to exactly one prediction - so duplicates never arise and no cleanup is needed.
On COCO, DETR reached accuracy comparable to a well-tuned Faster R-CNN while being conceptually simpler and requiring no specialized library, and the same model extended cleanly to panoptic segmentation. Its main weakness was slow convergence on small objects, which spawned a large follow-up literature (Deformable DETR and others).
Why business readers should care: DETR showed the transformer - the architecture behind modern language models - could replace decades of bespoke computer-vision engineering with a single, uniform design. That convergence on one architecture across vision and language is part of why multimodal models became practical to build.