DETR ResNet-50 (Object Det.)

DETR (DEtection TRansformer) ResNet-50 is an end-to-end object detection model from Meta that reimagines object detection as a direct set prediction problem. By combining a CNN backbone (ResNet-50) with a Transformer encoder-decoder, it eliminates the need for hand-designed components like anchor boxes and non-maximum suppression used in traditional detectors.

The model uses bipartite matching loss to directly predict a set of object bounding boxes and classes, simplifying the detection pipeline while achieving competitive accuracy. It handles multiple object categories and can detect objects at various scales within an image.

DETR represents a paradigm shift in object detection, demonstrating that Transformers can be effectively applied to vision tasks without the complex heuristics of traditional approaches.