Lillthorin/YOLOlite_edge_s

YOLOLite-edge_s (ONNX, 640x640)

YOLOLite-edge_s is a lightweight, CPU-oriented object detection model designed for real-time performance on edge devices and industrial systems.
The model is part of the YOLOLite family, focusing on practical latency, small model size, and stable performance across many real-world datasets.

---

📦 Full source code: https://github.com/Lillthorin/YoloLite-Official-Repo

📊 Benchmark Results can be found under BENCHMARK.md in the repo

🔍 Key Features

• Real-time CPU performance (25–30 FPS end-to-end on desktop CPU)
• Fast ONNX inference (14–18 ms per frame)
• Optimized for industrial and edge applications
• Supports resize or letterbox preprocessing
• Reliable performance across 40+ diverse Roboflow100 datasets

---

⚡ Real-World Performance (CPU, ONNX Runtime)

Tested using 1080p traffic footage (intersection.mp4) and the script
onnx_intersection_showcase.py.

Measurement	Result
End-to-end FPS	25–30 FPS
Raw ONNX inference	14–18 ms (~55–70 FPS)
Resolution	640 × 640
Execution Provider	CPUExecutionProvider

These values include the full pipeline: video → resize → inference → NMS → drawing.

---

🧪 Example Usage

from infer_onnx import ONNX_Predict
import cv2

predict = ONNX_Predict("edge_s_640_resize.onnx",
                       providers=["CPUExecutionProvider"],
                       use_letterbox=False)

frame = cv2.imread("image.jpg")
boxes, scores, classes = predict.infer_image(frame, img_size=640)

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