Rf-detr: advancing real-time transformer-based object detection model

Achieving an optimal trade-off among accuracy, inference speed, and cross-domain flexibility remains a central challenge in computer vision for real-time object detection. Conventional convolutional neural network (CNN) models support high throughput but lack contextual object understanding. Transformer-based detectors produce more accurate, context-aware results, yet they require prolonged convergence and incur high computational cost, rendering them unsuitable for real-time deployment. To address these limitations, this paper proposes the Real-Time Flexible Detection Transformer (RF-DETR), a hybrid object detection architecture that integrates recurrent feature refinement, flexible deformable attention, and CNN-based multi-scale feature extraction. RF-DETR achieves high detection accuracy and low inference latency without requiring post-processing steps such as Non-Maximum Suppression. The model is evaluated on two benchmark datasets: COCO 2017 for primary training and testing, and RF100-VL for cross-domain evaluation. Experimental results demonstrate that RF-DETR surpasses representative CNN-based and transformer-based detectors in mean average precision (mAP), recall, and inference time. With a task accuracy of 94.85% and real-time performance on both GPU and edge-compatible hardware, RF-DETR constitutes a robust and scalable solution for challenging, time-sensitive object detection applications.