Parameters10.0M
GFLOPs25.0
Input Size640px
Best mAP53.5%
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 | 50.7% | 24.0 | 41.7ms | 140 MB |
| NVIDIA Jetson Orin Nano Super 8GB | ONNX Runtime FP32 | 53.4% | 10.4 | 96.1ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | PyTorch FP32 | 53.4% | 10.3 | 96.8ms | 139 MB |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP16 | 53.5% | 30.2 | 33.1ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP32 | 53.4% | 22.4 | 44.7ms | - |
| NVIDIA RTX 5070 Ti | ONNX Runtime FP32 | 53.4% | 70.9 | 14.1ms | - |
| NVIDIA RTX 5070 Ti | PyTorch FP32 | 53.4% | 34.2 | 29.3ms | 142 MB |
| NVIDIA RTX 5070 Ti | TensorRT FP16 | 53.2% | 86.5 | 11.6ms | - |
| NVIDIA RTX 5070 Ti | TensorRT FP32 | 53.4% | 98.7 | 10.1ms | - |
Speed Breakdown(NVIDIA A100)
6.2ms
23.8ms
11.8ms
Preprocess
Inference
Postprocess (NMS)
Usage with LibreYOLO
from libreyolo import LibreYOLO
# Load model (auto-downloads from HuggingFace if not found locally)
model = LibreYOLO("LibreDFINEs.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: 48.7% mAP