This is a brief overview on the project, there is pdf of a presentation in /project folder. As part of the course we were required to find a project and implement it on a microcontroller. Most projects on a microcontroller involve a CNN, to bring novelty but still leverage the architecture of the MAXIMM78000 I chose to implement a U-Network. My idea was to be able to segment a face from a picture and then use the segmented face to do some other task. For example, I could use the segmented face to do face recognition or to do emotion recognition. For the training of the U-Net I used the CelebAMask-HQ Dataset. Since the maximum input size of the network is 90×91x3, I created a dataloader to make the image to be 88x88x3.
Here are a few results of the segmentation:
These are the results:
| Data-Set | Size | IoU |
|---|---|---|
| Training | 22746 | 0.91390 |
| Test | 4836 | 0.88632 |
The main things that were done are:
- adding a dataloader to the project
- changing the existing U-Net to fit on the MC
- adding a new loss function (Dice Loss)
- adding the status of the training
- adding a new metric (IoU)
- training the network
- Synthesizing the network using the MAXIM toolchain.
- Testing the network on the MAXIM board
The project is not finished, there are still a few things that were started but could not be verified due to time constraints:
- adding other things to segment (like eyes, mouth, etc.)
- adding a loss function for the other things to segment
Details on the exercises can be found in /exercises or on the courses website