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:
The annotated samples like the below:
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
The MobileNetV1-FPN-SSD architecture may like this below:
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.
| Method | mAP | mAP(large) | mAP(medium) | mAP(small) | mAP@0.50IOU | mAP@0.75IOU |
|---|---|---|---|---|---|---|
| MobileNet-SSD | 45.20 | 62.58 | 48.14 | 13.14 | 75.79 | 45.81 |
| MobileNet-SSD-FPN | 60.01 | 75.03 | 64.46 | 29.25 | 84.65 | 71.85 |
