Parameters4.0M
GFLOPs7.0
Input Size640px
Best mAP45.8%
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 | 42.8% | 25.6 | 39.1ms | 68 MB |
| NVIDIA Jetson Orin Nano Super 8GB | ONNX Runtime FP32 | 45.8% | 17.5 | 57.0ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | PyTorch FP32 | 45.8% | 11.7 | 85.5ms | 64 MB |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP16 | 45.8% | 42.1 | 23.8ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP32 | 45.8% | 33.5 | 29.9ms | - |
| NVIDIA RTX 5070 Ti | ONNX Runtime FP32 | 45.8% | 78.9 | 12.7ms | - |
| NVIDIA RTX 5070 Ti | PyTorch FP32 | 45.8% | 32.5 | 30.7ms | 70 MB |
| NVIDIA RTX 5070 Ti | TensorRT FP16 | 45.8% | 79.7 | 12.5ms | - |
| NVIDIA RTX 5070 Ti | TensorRT FP32 | 45.8% | 98.7 | 10.1ms | - |
Speed Breakdown(NVIDIA A100)
6.1ms
21.3ms
11.6ms
Preprocess
Inference
Postprocess (NMS)
Usage with LibreYOLO
from libreyolo import LibreYOLO
# Load model (auto-downloads from HuggingFace if not found locally)
model = LibreYOLO("LibreDFINEn.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: 42.8% mAP