Most of my effort in the week before fall break was spent on the design report, with midterms for other classes taking up much of my time as I expected. Most of the contents of the design report was laying out our existing plan for the system, with one exception being that we are considering a web application component for showing recycling statistics as an add-on to the display system that was described. However, the CV system (my concentration) is pretty much unchanged from the design report, so no major changes are planned.
I was unexpectedly busy studying for interviews over fall break, but I did have time to work on setting up the Jetson and starting to train the model. Setting up the Jetson was more involved than I anticipated. I did have a microSD card on hand to hold the Jetson Orin Nano SDK image, but was having trouble getting anything to show when connecting the Jetson to a display. Looking at the documentation, I believe the issue is that Jetson only supports connecting to external displays via the DisplayPort port, so next week I will try booting the Jetson with a DisplayPort cable, and hopefully explore the Jetson software. With regards to model training, my original plan was to train YOLOv7 using the Jetson’s GPU, but Google Colab will work as well, even with usage limits. The original dataset I was looking at, Drinking Waste Classification, already comes with the annotation format that YOLO expects for custom datasets, so I will starting with that. Next week I will partition the dataset (it was not partitioned unfortunately), and run training. From there, I will tweak training parameters and evaluate the model’s performance (without the Jetson).
I am also looking into other datasets, since I realized the drinking waste dataset doesn’t really cover all recyclable materials (ie. cardboard is not included). I’ve also found some datasets of images of electronic items (https://www.kaggle.com/datasets/dataclusterlabs/electronics-mouse-keyboard-image-dataset, https://www.kaggle.com/datasets/ksenia5/electronic-object-detection), which could be used to train our model to recognize special waste that can’t be disposed in the trash.