Parameters31.0M
GFLOPs91.0
Input Size640px
Best mAP52.3%
LicenseApache-2.0
Architecture
Type
transformer
Backbone
ResNet-34
Neck
HybridEncoder
Head
DETR
Benchmark Results
Performance on COCO val2017 across different hardware configurations
| Hardware | Runtime | mAP@50-95 | FPS | Latency | VRAM |
|---|---|---|---|---|---|
| NVIDIA A100 | PyTorch FP32 | 48.2% | 27.8 | 36.0ms | 198 MB |
| NVIDIA Jetson Orin Nano Super 8GB | ONNX Runtime FP32 | 52.2% | 7.6 | 131.7ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | PyTorch FP32 | 52.2% | 8.0 | 125.8ms | 198 MB |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP16 | 52.3% | 31.9 | 31.4ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP32 | 52.2% | 17.1 | 58.5ms | - |
| NVIDIA RTX 5070 Ti | ONNX Runtime FP32 | 52.2% | 77.7 | 12.9ms | - |
| NVIDIA RTX 5070 Ti | PyTorch FP32 | 52.2% | 40.2 | 24.9ms | 213 MB |
| NVIDIA RTX 5070 Ti | TensorRT FP16 | 52.3% | 108.4 | 9.2ms | - |
| NVIDIA RTX 5070 Ti | TensorRT FP32 | 52.2% | 98.1 | 10.2ms | - |
Speed Breakdown(NVIDIA A100)
3.2ms
23.9ms
9.0ms
Preprocess
Inference
Postprocess (NMS)
Usage with LibreYOLO
from libreyolo import LibreYOLO
# Load model (auto-downloads from HuggingFace if not found locally)
model = LibreYOLO("LibreRTDETRr34.pt")
# Run inference
result = model("image.jpg", conf=0.25, iou=0.45)
# Process results
print(f"Found {len(result)} objects")
print(result.boxes.xyxy) # bounding boxes (N, 4)
print(result.boxes.conf) # confidence scores (N,)
print(result.boxes.cls) # class IDs (N,)detrnms-freePaper: 48.5% mAP