This repo is similar to the handtracking. I train the nail detector using MobileNetV1-FPN-SSD, and the dataset I used is annotated by laboratory member.

Here is the nail detector in action and draw the line:

nail_draw

The annotated samples like the below:

1560172595651

The overall process is divided into 3 steps:

• Histogram equalization

• Skin color binarization(convert the RGB channel to YCbCr and use inRange method to binarize)

• MobileNet-SSD-FPN

  1560172217358

The MobileNetV1-FPN-SSD architecture may like this below:

ssd_fpn_nail

I use the MobileNet-SSD first, but the nail is too tiny to detect. So we can see the MobileNet-SSD-FPN is pretty better than MobileNet-SSD.

1560173400082