RF-DETR

RF-DETR optimizes end-to-end object detection by utilizing Receptive Field-based feature fusion to eliminate redundant computations in Transformer encoders.

RF-DETR addresses the high computational cost of the DETR architecture by introducing a Receptive Field-augmented encoder. By focusing on multi-scale feature fusion and spatial constraints, this model reduces the complexity of global self-attention mechanisms without sacrificing mAP scores. The implementation leverages a lightweight encoder design that achieves competitive performance on the COCO dataset (reaching 44.6 mAP with a ResNet-50 backbone) while significantly lowering GFLOPs. It is a precise solution for developers needing the accuracy of query-based detection on hardware with constrained processing power.

https://github.com/SlongLiu/RF-DETR