What did you personally accomplish this week on the project?
- First I did more substantial work on the Grounded Dino implementation. Essentially I confirmed that Grounded Dino works to a reasonable degree in many respects it matches the results that I got with my training with YOLO. In order to get just how much better I began working on a python script to compare outputs for Grounded DINO and YOLO.
- After completing training on the latest model on the 30th I noticed that for some reason there was a major issue generalizing to do different types of pencils/pens. That is because the existing training data used to build up the other existing categorizations in the model has pencils/pens inside of it. All the false negatives are probably not enabling the model to actually see any different types of pencils as pencils. I’m currently testing approaches ranging from manual to GPT/Grounded Dino to manually identify images with pencils in them so I can then label them manually. There are only 300 of those images out of like 50,000 so the time consuming part is finding these images in the first place.
- One major breakthrough is that I began testing my models on the raspberry Pi. The initial results are looking good enough for my YOLO approach though further work that can easily be displayed has to be done.
- I have fully 3D printed and assembled parts related to the rasp pi box.
Is your progress on schedule or behind? If you are behind, what actions will be
taken to catch up to the project schedule?
My progress is somewhat on schedule as I am getting increasingly good vision algorithms though I would argue that the other non-critical/not very hard work like audio processing is not done. At some point in the next 2 weeks I’ll just have to accept the model that I have at the time and work on the other components.
What deliverables do you hope to complete in the next week?
I am aiming to release graphics comparisons of the YOLO vs Grounded DINO and then perhaps hybrid approaches.
I would also like to complete testing on the rasp pi to feel satisfied that everything works as expected on its camera and is ready for demo. I also hope to figure out the issue relating to correctly labelling pencils/pens/other add ons in the original dataset i got from online.
If time permits I would also like to do more work on the audio processing/vector comparison work
0 Comments