4 ms·
Hardest part of building computer vision models these days is collecting and annotating data. A self serve tool like sagemaker doesn't solve that problem. As a
by m_ke 7y ago
Hardest part of building computer vision models these days is collecting and annotating data. A self serve tool like sagemaker doesn't solve that problem.
As a computer vision company your best bet is to focus on a specific industry or a narrow problem and build fully featured products. This could be medical imaging, satelite imagery, face recognition, fashion, adult content filtering, face recognition, etc.
We're doing this for food. We built a consumer food logging app that's a great source of data to keep improving our recognition model. At this point we have millions of labeled images, something that would take most of our API customers years to do. To make integration easier we couple the recognition model with a food "knowledge graph" that includes nutrition facts, ingredients and dietary categorizations. It turns out that this is useful to a lot of healthcare, fitness, nutrition and CPG companies.
- grizzly_bear 7y agoYou're right, this is a vertical approach. Some of our customers are companies like this, having a focused model for a very specific subject.
- codingslave 7y agoYeah, I'm working on a computer vision product, and the dtaa collection is a nightmare. Not just collecting alot, but getting enought samples for each class and each type of situation that might arise. It makes me want to pull my hair out.
- grizzly_bear 7y agoTricky question: how do you define "enough"? :) (I usually answer my clients as a joke: "thing about the MAXIMUM number of images you can provide me. Well 'enough', for us, is twice more!")
- codingslave 7y agoWhen I've tuned my models a million ways and they still overfit? I'm suspecting I don't have enough
- grizzly_bear 7y agoI mean, the tricky part is to have an estimation of the number of images before you even train your model :) (usually for budgeting reasons)