Parameters18.0M
GFLOPs53.0
Input Size640px
Best mAP58.4%
LicenseApache-2.0
Architecture
Type
transformer
Backbone
ECViT (compact ViT)
Neck
HybridEncoder
Head
DETR
Benchmark Results
Performance on COCO val2017 across different hardware configurations
| Hardware | Runtime | mAP@50-95 | FPS | Latency | VRAM |
|---|---|---|---|---|---|
| NVIDIA Jetson Orin Nano Super 8GB | ONNX Runtime FP32 | 58.4% | 3.9 | 254.4ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | PyTorch FP32 | 58.4% | 5.0 | 200.8ms | 160 MB |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP16 | 58.3% | 8.3 | 119.7ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP32 | 58.3% | 8.2 | 121.6ms | - |
| NVIDIA RTX 5070 Ti | ONNX Runtime FP32 | 58.4% | 40.2 | 24.9ms | - |
| NVIDIA RTX 5070 Ti | PyTorch FP32 | 58.4% | 26.5 | 37.7ms | 162 MB |
| NVIDIA RTX 5070 Ti | TensorRT FP16 | 58.2% | 48.9 | 20.5ms | - |
| NVIDIA RTX 5070 Ti | TensorRT FP32 | 58.4% | 59.2 | 16.9ms | - |
Speed Breakdown(NVIDIA Jetson Orin Nano Super 8GB)
16.3ms
181.5ms
3.0ms
Preprocess
Inference
Postprocess (NMS)
Usage with LibreYOLO
from libreyolo import LibreYOLO
# Load model (auto-downloads from HuggingFace if not found locally)
model = LibreYOLO("LibreECm.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