Based on only about 120 images from open datasets and my own generation. Therefore, it should get better results on realistic ones.

Threshold set to 0.4~0.7 might be good.