Skip to main content
This page reports RF-DETR benchmark results for object detection and instance segmentation on Microsoft COCO and RF100-VL. All benchmark numbers match the latest released checkpoints. For full methodology details and architectural context, see the RF-DETR arXiv paper.

Methodology

Accuracy is reported using standard COCO metrics computed with pycocotools. For object detection, we report COCO AP50 and COCO AP50:95, and the same metrics are also reported for RF100-VL. COCO results are evaluated on the validation split, following common practice in detector benchmarking.
RF100-VL results are averaged across 100 diverse real-world datasets to reflect performance under diverse data distributions — making it a strong indicator of how a model generalises beyond COCO.
Latency is measured as single-image inference latency rather than sustained throughput. All latency numbers are obtained on an NVIDIA T4 GPU using TensorRT 10.4 and CUDA 12.4 with FP16 inference and batch size 1. To reduce variance caused by GPU power throttling and thermal effects, a 200 ms buffer is inserted between consecutive forward passes. This procedure improves reproducibility of latency measurements but is not intended to measure maximum throughput.
Accuracy and latency are always measured using the same model artifact and the same numerical precision. This avoids reporting FP32 accuracy with FP16 latency, which can lead to misleading comparisons because naive FP16 conversion can significantly degrade accuracy for some models.

Results

Benchmarks compare RF-DETR against YOLO11, YOLO26, LW-DETR, and D-FINE on Microsoft COCO and RF100-VL.
ArchitectureCOCO AP50COCO AP50:95RF100VL AP50RF100VL AP50:95Latency (ms)Params (M)Resolution
RF-DETR-N67.648.485.057.72.330.5384x384
RF-DETR-S72.153.086.760.23.532.1512x512
RF-DETR-M73.654.787.461.24.433.7576x576
RF-DETR-L75.156.588.262.26.833.9704x704
RF-DETR-XL77.458.688.562.911.5126.4700x700
RF-DETR-2XL78.560.189.063.217.2126.9880x880
YOLO11-N52.037.481.455.32.52.6640x640
YOLO11-S59.744.482.356.23.29.4640x640
YOLO11-M64.148.682.556.55.120.1640x640
YOLO11-L64.949.982.256.56.525.3640x640
YOLO11-X66.150.981.756.210.556.9640x640
YOLO26-N55.840.376.752.01.72.6640x640
YOLO26-S64.347.782.757.02.69.4640x640
YOLO26-M69.752.584.458.74.420.1640x640
YOLO26-L71.154.185.059.35.725.3640x640
YOLO26-X74.056.985.660.09.656.9640x640
LW-DETR-T60.742.984.757.11.912.1640x640
LW-DETR-S66.848.085.057.42.614.6640x640
LW-DETR-M72.052.686.859.84.428.2640x640
LW-DETR-L74.656.187.461.56.946.8640x640
LW-DETR-X76.958.387.962.113.0118.0640x640
D-FINE-N60.242.784.458.22.13.8640x640
D-FINE-S67.650.685.360.33.510.2640x640
D-FINE-M72.655.085.560.65.419.2640x640
D-FINE-L74.957.286.461.67.531.0640x640
D-FINE-X76.859.386.962.211.562.0640x640

Build docs developers (and LLMs) love