arthors/carwin-element-detection

Carwin Desktop UI Element Detection (YOLOv5n)

A lightweight YOLOv5n model trained to detect interactive UI elements on desktop screens (buttons, icons, input fields, checkboxes, etc.).

Model Details

• Architecture: YOLOv5n (ReLU activation, SiLU→ReLU for RKNN/NCNN compatibility)
• Input: 640×640 RGB
• Output: Single class "element" — bounding boxes for interactive UI regions
• Model size: 5.3 MB (PyTorch), 7.5 MB (ONNX)
• Training: 200 epochs on NVIDIA B200 (single GPU)

Performance

Metric	Value
mAP@0.5	0.925
mAP@0.5:0.95	0.648
Training images	10,825

Edge Deployment

Deployed to ARM Cortex-A7 via NCNN INT8 quantization:

Resolution	Inference Time
640×640	6.5s
320×320	1.6s
160×160	370ms

ONNX → NCNN → INT8 quantization pipeline included in the training repository.

Usage

import torch

# Load model
model = torch.hub.load('ultralytics/yolov5', 'custom', path='model.pt')
model.conf = 0.25

# Run inference
results = model('screenshot.png')
results.show()

Files

• model.pt — PyTorch weights (5.3 MB)
• model.onnx — ONNX export (7.5 MB, opset 12, batch=1, 640×640)

Training

Trained from yolov5n.pt pretrained weights with:

python train.py --data dataset.yaml --weights yolov5n.pt \
    --epochs 200 --batch-size 64 --device 0 --imgsz 640 \
    --single-cls --amp=False

• GPU: NVIDIA B200 (180 GB)
• PyTorch: 2.12 + CUDA 13.0
• Dataset: 10,825 annotated desktop screenshots

License

MIT