I've developed a mobile app using Google Vertex AI and its default model, which works decently well for my specific needs. However, I've noticed that apps like Plant ID, Chicken ID, and CoinID seem to generate highly specific results, possibly due to custom models. For instance, the Plant ID app assesses the health of plants based on leaf appearance, and the Chicken ID app likely tracks genetics. Given that I have over 40,000 testing images across 125 labels, it raises the question of whether these popular apps are built on public models or if they invest in custom development. Do these developers have significant funding to hire professional developers for their models? If there are any other relevant communities to explore this, please let me know!
2 Answers
Many of those apps were developed long before public models were available. They likely created custom models from scratch, which involved years of effort and huge datasets. For example, the Plant ID app originally had about 300,000 images! They're continually updating it too.
I don’t know the specifics for those apps, but another approach is to take an existing model and perform additional training on it, which helps adapt the model to your specific requirements.
What does 'tuning it' actually involve?
So, if they have that many images, do they just create several datasets if Vertex Object Detection only allows 150k images? Would they run tests on each and then choose the best results?