This week I was able to get YOLOv7 up and running on my computer to test its performance on some sample images. This was purely to gauge the model out of the box, and as expected, the classification was not very good on recyclable materials.
This was expected, and confirms our need to find a suitable dataset of recyclable materials to fine tune the model. Our original plan was to train our model on some labeled recyclable materials (ie. aluminum can, plastic bottle), and have some sort of lookup table for each classification to determine if the item is recyclable. This could potentially lead to cases where an item isn’t recognized with high confidence, and we would have to make a decision on what to do in such a case. An alternative we discussed was customizing our dataset so items are labeled as “recyclable” or “non-recyclable”, but currently I am leaning towards the original idea, with unrecognized items being simply classified as non-recyclable.
Aside from CV work, most of my time was spent on the design presentation, and working out the details of our new game system. The game system won’t interact directly with the CV, but adding a new system to the project felt like something for the whole group to discuss.
Progress has been good so far. I will be (temporarily) leaving the country in the middle of next week. I should (and plan to), work on assignments while I am gone, and luckily I should be able to do CV work remotely, but between being gone, and making sure that I’m keeping up with other classes. I’m anticipating that I may not get as much work done next week. I’m hoping to find a suitable dataset and figure out how to train the existing model on this dataset. One that I have found is this: https://www.kaggle.com/datasets/arkadiyhacks/drinking-waste-classification , but plenty more are available, and I’d like to find a dataset that may include items that are often thought to be recyclable but actually aren’t (ie. plastic bags). If the Jetson camera arrives before I leave, I also plan to hook it up to get the Jetson to capture images.
I will make sure to communicate with my group if I have any issues.