Back to Leaderboard

RT-DETR-R18

rtdetr

transformer detector with ResNet-18 backbone

Parameters

20.0M

FLOPs

60.0G

Input Size

640px

License

Apache-2.0

Architecture

Type

transformer

Backbone

ResNet-18

Neck

HybridEncoder

Head

DETR

Benchmark Results
Performance on COCO val2017 across different hardware configurations
HardwaremAP@50-95FPSLatencyVRAM
NVIDIA A100 (TensorRT FP16)46.5%114.98.7ms914 MB
NVIDIA T4 (TensorRT FP16)46.5%45.621.9ms868 MB
CPU (ONNX Runtime)46.5%5.3187.9ms996 MB
Speed Breakdown (A100 TensorRT)
End-to-end latency breakdown showing preprocessing, inference, and postprocessing times
1.1ms
4.9ms
2.7ms
Preprocess
Inference
Postprocess (NMS)
Usage with LibreYOLO
from libreyolo import YOLO

# Load model
model = YOLO.from_pretrained("https://huggingface.co/Libre-YOLO/rtdetr-r18")

# Run inference
results = model.predict("image.jpg")

# Process results
for box in results.boxes:
    print(f"Class: {box.cls}, Confidence: {box.conf:.2f}")
transformerno-nms
Notes

First real-time DETR, no NMS required