Parameters76.0M
GFLOPs259.0
Input Size640px
Best mAP56.8%
LicenseApache-2.0
Architecture
Type
transformer
Backbone
ResNet-101
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 | 53.9% | 18.6 | 53.6ms | 508 MB |
| NVIDIA Jetson Orin Nano Super 8GB | ONNX Runtime FP32 | 56.8% | 2.9 | 343.9ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | PyTorch FP32 | 56.8% | 2.8 | 359.6ms | 450 MB |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP16 | 56.8% | 17.0 | 58.8ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP32 | 56.8% | 8.4 | 118.5ms | - |
| NVIDIA RTX 5070 Ti | ONNX Runtime FP32 | 56.8% | 47.5 | 21.1ms | - |
| NVIDIA RTX 5070 Ti | PyTorch FP32 | 56.8% | 23.6 | 42.4ms | 456 MB |
| NVIDIA RTX 5070 Ti | TensorRT FP16 | 56.7% | 88.0 | 11.4ms | - |
| NVIDIA RTX 5070 Ti | TensorRT FP32 | 56.8% | 64.4 | 15.5ms | - |
Speed Breakdown(NVIDIA A100)
3.2ms
42.1ms
8.3ms
Preprocess
Inference
Postprocess (NMS)
Usage with LibreYOLO
from libreyolo import LibreYOLO
# Load model (auto-downloads from HuggingFace if not found locally)
model = LibreYOLO("LibreRTDETRr101.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: 54.3% mAP