README.md

---
license: apache-2.0
---

# YOLOLite-edge_s (ONNX, 320×320, P2 head)

YOLOLite-edge_s is a lightweight, CPU-focused object detection model designed for **extreme edge performance** at very low resolutions.  
This P2-enabled variant is optimized for **small-object detection at 320×320** while maintaining high real-time throughput on ordinary CPUs.

📦 Full source code: https://github.com/Lillthorin/YoloLite-Official-Repo  
📊 Full Benchmark Results: See `BENCHMARK.md` in the repository

---

## 🔍 Key Features

- **Real-time CPU throughput: 94–101 FPS end-to-end**  
- **Fast ONNX inference: 8–10 ms per frame**  
- Optimized for **industrial, robotics, and edge computing**  
- Enhanced **P2 head** for small-object performance at 320px  
- Supports **resize** or **letterbox** preprocessing  
- Evaluated across **40+ diverse Roboflow100 datasets**

---

## ⚡ Real-World CPU Performance (ONNX Runtime)

Tested on 1080p traffic footage (`intersection.mp4`) using  
`onnx_intersection_showcase.py` with:

- Model: `edge_s_320_p2.onnx`
- Execution Provider: `CPUExecutionProvider`
- Preprocessing: Resize
- Resolution: **320×320**

| Measurement | Result |
|------------|--------|
| **End-to-end FPS** | **94–101 FPS** |
| **Raw inference latency** | **8–10 ms** per frame |
| **Pipeline includes** | video → resize → inference → NMS → drawing |

These values represent **actual full-pipeline performance**, not isolated model latency.

---

## 🧪 Example Usage

```python
from infer_onnx import ONNX_Predict
import cv2

predict = ONNX_Predict(
    "edge_s_320_p2.onnx",
    providers=["CPUExecutionProvider"],
    use_letterbox=False
)

frame = cv2.imread("image.jpg")
boxes, scores, classes = predict.infer_image(frame, img_size=320)

for (x1, y1, x2, y2), score, cls in zip(boxes, scores, classes):
    print(x1, y1, x2, y2, score, cls)