I developed a Landmark program based on the simple UNet, which I have shared: https://github.com/Achillesy/CL-DETECTION2023

The program can be run directly on colab. At low resolution (256x256) without image augmentation, one landmark result can be trained in a very short time (within 10 minutes).

Some points are not bad, like name=26 (cfg.name=25); however, when I train name=27 (cfg.name=26), I find it harder than other points. I think the output Mask is as expected, but the prediction result is not ideal.

So I would like to ask what exactly the name=27 tag indicates, and what are the features around this?