Parameters31.0M
GFLOPs91.0
Input Size640px
Best mAP60.0%
LicenseApache-2.0
Architecture
Type
transformer
Backbone
HGNetv2
Neck
HybridEncoder
Head
FDR (Fine-grained Distribution Refinement)
Benchmark Results
Performance on COCO val2017 across different hardware configurations
| Hardware | Runtime | mAP@50-95 | FPS | Latency | VRAM |
|---|---|---|---|---|---|
| NVIDIA A100 | PyTorch FP32 | 57.3% | 17.1 | 58.3ms | 245 MB |
| NVIDIA Jetson Orin Nano Super 8GB | ONNX Runtime FP32 | 60.0% | 5.0 | 199.6ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | PyTorch FP32 | 60.0% | 4.7 | 211.7ms | 246 MB |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP16 | 60.0% | 18.6 | 53.6ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP32 | 60.0% | 12.5 | 80.2ms | - |
| NVIDIA RTX 5070 Ti | ONNX Runtime FP32 | 60.0% | 50.9 | 19.6ms | - |
| NVIDIA RTX 5070 Ti | PyTorch FP32 | 60.0% | 21.9 | 45.8ms | 245 MB |
| NVIDIA RTX 5070 Ti | TensorRT FP16 | 60.0% | 69.3 | 14.4ms | - |
| NVIDIA RTX 5070 Ti | TensorRT FP32 | 60.0% | 72.3 | 13.8ms | - |
Speed Breakdown(NVIDIA A100)
6.3ms
41.4ms
10.7ms
Preprocess
Inference
Postprocess (NMS)
Usage with LibreYOLO
from libreyolo import LibreYOLO
# Load model (auto-downloads from HuggingFace if not found locally)
model = LibreYOLO("LibreDFINEl.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% mAP