Hello! I am a highschool student working on a project to create a drone that autonomously detect and pick up plastic water bottles. Currently, the limit for one project is 10,000 images, but to train the AI to become more accurate, I would need more images. Would this be possible without paying?
2 Likes
We support students with expanded credits. Have you applied for research credits: Contribute to the Computer Vision Community
1 Like
Hey Mahin_Gupta! Sounds like a cool project! Roboflow will be a great way to get the vision piece up and running. You may have already seen some of this as you have researched, but one thought for you - start small.
Just annotate 50 images, do some augmentations and then train a model. It’ll be pretty good. If you want to improve, do a few hundred. And worst case, I do not expect you to need more than maybe 1,000 for a viable project. Roboflow has a nice article on how many images companies tend to use - Computer Vision Trends Report 2025 - Key Benchmarks (Keep in mind that the few companies using thousands of images likely have some major manufacturing processes that require high accuracy for critical operations. You might not need to be that robust.)
Have fun building that out!
1 Like
I’ve actually tried with fewer images, however since the camera feed is live and constant, it will sometimes pick up objects that are not water bottles or will not get water bottles from far distances.
2 Likes
Hi @Mahin_Gupta,
This is an incredibly exciting project and is similar to another project I worked on in the past.
You may have come across this in your research, but to reiterate, ensure that your training set includes edge cases such as objects commonly mistaken for water bottles and objects at various distances to help minimize false positives. Also, fine-tuning the anchor box sizes and confidence thresholds can further help the model reduce false positives.
For example, here is a high level overview of a workflow that you could implement. Using the Roboflow Dataset Upload allows you to quickly review predictions, accept or reject them, and add to your training set with new labeled examples.
Additionally, to better handle the constant camera feed you can use frame sampling to limit redundant inferences.
Please keep us updated on your progress!
1 Like
This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.