It’s probably a trained CNN underneath then. In theory, I could build a CNN like that. The hard part to making it work well is gathering and accurately labeling a large set of training data. Their website is picturethisai.com, ai kind of pointing me to them using some type of (or multiple) neural nets.
They don’t disclose their process but they mention it doesn’t work offline because it uses a huge database. A trained neural network doesn’t need a huge database anymore so that makes me think they’re trying to obfuscate the process and/or they have a multi step algorithm that may rely on some traditional computer vision techniques.
If you’re starting from scratch, I’d say look into tensorflow for python.
A CNN is a convolutional neural net. The convolutions being matrices multiplication (linear algebra) which is a traditional computer vision filtering technique. However, CNNs also backwards propagate weight adjustments which is where the chain rule (algebra/calculus) come into play.
The state of AI development libraries is such that you don’t need to understand the math or even the techniques themselves to build and utilize them. Which is a double edged sword but that’s also how technology progresses.
To relate CNNs to something we all have been exposed to at least via the news and media, the Tesla auto pilot uses a trained segmentation CNN. It tries to label every pixel in a camera frame as some object or another. There’s probably a lot of videos demonstrating the output of that type of network.
Back in college I took one ML class. it was so boring, all I was doing was implementing neural networks. What are the problems with AI Development. What makes a company truly an AI company? Are they building their own AI or are they using other peoples AI?
So if it actually said “this plant has bugs” or “this plant has mites” it would be much more helpful than “this is marijuana it looks under watered etc”
So I built an application for a company that uses eyetracking to measure visual fixation/interest. It’s currently being used as an accepted form of 3rd party testing for use in courts related to sex offenders. Basically show a grid of images and measure how many times and for how long the person looks at the various defined regions of interest. It was a neat project that took me a couple years but it’s wild that the results of my code is being used in court cases to help prevent predators from sliding through the broken system.
All python baby. PyQt for gui, numpy, firebase, Tobii for eyetracking, and CherryPi for the server front end. I even rolled my own encryption for the data collected if for some reason it lost internet connection, it would encrypt and save as file for later upload.
It was a client/server model where the data acquisition happened on the client and the data was sent to the server for processing. Managed to get 13 million data points in under 15 minutes. Also managed to get them processed in under 30s at O(n^2). Funnily enough I even managed to develop a linear algorithm for it but due to all of the function calls I had to make, it took longer to run my linear function than it did to run it at O(n^2).
