What are the most significant risks that could jeopardize the success of the project? How are these risks being managed? What contingency plans are ready?
The most significant risk as of now is to ensure that we are able to deploy our model on the board we choose. For a raspberry pi + AI kit, using an external accelerator and connecting that to our main processor will be a tedious process, and from previous experience there is a lot of debugging involved. As for the Nvidia Jetson Nano, it will be much more convenient since the GPUs are included in the kit, but none of us have experience with using the board, so it’ll be a challenge to ensure that we can get the board up and running and have all the software stack set up. Another challenge is to ensure that all of our input/output data processing can be done. Once we decide on the camera and the LCD displays, writing the scripts for image processing (to feed into our model) and producing output data (analysis on each of the classes) can be challenging, since we do not have embedded experience with these boards. We are currently looking into 2 different boards with compatible cameras/lcd displays so that we have an alternative plan. However, because the parts are expensive, we want our first choice to work and are doing extensive research in open source CV projects that work with the boards we researched.
Another potential risk is the classification accuracy on the oiliness detection task. Unlike the other tracked factors, oiliness is much less visually clear. This could potentially result in a sub-85% classification error when trained only on the initial datasets. To counteract this, if model performance is below our expectations, we will gather new data points through web-scraping, and potentially new images we collect ourselves. Models to detect oiliness have been implemented prior, so likely gathering more data will drive down the error rates.
The datasets available for skin burns are either locked behind stock image distributors paywalls, or contain many images of much greater severity, and also don’t solely focus on the face. Potentially, this could cause a bias in our sunburn detection model to only classify significant burns. To counteract this, we will augment the existing dataset we were using with images of sunburns scraped from google images. Google image results for sunburns are much better aligned with the goals of our system, and match typical levels of sunburn people face in their day-to-day life.
Were any changes made to the existing design of the system (requirements, block diagram, system spec, etc)? Why was this change necessary, what costs does the change incur, and how will these costs be mitigated going forward?
We changed our computer vision model backbone from MobileNetV2 to MobileNetV3-small. MobileNetV3 is a more modern iteration, and has better performance on edge devices, while retaining similar classification accuracy (defined on imagenet top-1 and top-5 classification rate).
Upon further research, our team is now considering removing the two MCU units in our hardware block diagram. The goal of the project can be achieved with just a single computer board directly interfacing with our camera inputs and LCD outputs. This adjustment simplifies the architecture itself, but the complexity of the project remains as we are still interfacing with the same hardware components from the plan.
This design change, however, requires corresponding updates to our communication protocol and power supply to ensure compatibility with the connected devices. In addition, if we choose to work with a Raspberry Pi 5, we will have an external AI accelerator to meet our use-case requirement of displaying results on the LCD within the 7 seconds from the user being positioned correctly.
Provide an updated schedule if changes have occurred.
This week, we originally intended to finalize selecting the different hardware components for our project. However, upon research, we realized that a discussion with our advisors will be needed in order to make a decision and compile a parts list. This influenced our search for libraries as they will depend on which single board computer we select for our project. We will be catching up to our schedule until Wednesday of Week 5.