Personal Image Classifier does not work

Hello guys, as every Sunday I dedicate some time to learn interesting things... I am learning how to implement image calssification using AI2 and the personal image classifier extension. My problem is that I upload a set of 5 images to create the model using https://classifier.appinventor.mit.edu/oldpic/ or https://classifier.appinventor.mit.edu, then I select the model and press Train... and then... nothing, the time does not increment in both classifier web tools. I do not think it is normal, I did not see any activity in more than 30 minutes. What I am missing or what I am doing wrong?

Maybe a community search can provide an answer?

https://community.appinventor.mit.edu/search?q=personal%20image%20classifier

1 Like

I did but nothing helped me. In practice yesterday nothing worked, I left it all the night but there were no progress at all and I loaded only 5 images, so I would understand if the size of the images is a problem but both versions did not signal any error, they just did not show any progress. The UI is fairly simple so...

or

might help.

1 Like

Thanks, I already saw them and followed their instructions... yesterday did work nothing... today seems that at least with little test images the old version works. I have a couple of meetings today, I hope to repeat my tests this evening but anyway the situation is better then yesterday. May were related to internet bandwidth? Yesterday there were some thunderstorms in my area and not so far and yes... in this condition we have problems; but anyway I did not receive any error or warning from the 2 PIC online tools. P.S. for my tests I am using about 50 pictures taken around my house (mainly paintings and about 5 pictures for each one, each picture taken with my smartphone is about 1.2MB) then I would like to realize an app similar to PICaboo able to recognize the paintings and display few info about them. It is just a quick experiment.

What are the hardware specifications for your computer? Training PIC requires that your browser and hardware provide support for WebGL 2.0. If the progress bar is not progressing, the software is likely failing to initialize correctly due to the hardware not supporting it.

Now everything is OK, probably it was an internet related problem in this area. Instead I am obtaining strange results, I would need more info about it works. Until I instruct the neural network to distinguish between 2 different situations (as may be face with or without glasses) it works reliably but if I want to distinguish among more categories it fails with a lot of errors. My second try was to check if it is able to distinguish among some paintings, a plate, a poster, a painting on the wall, a picture of a seaside, an image of Donald Duck and a chinese hat. In this case the results are bad. I am interested to understand if it may recognize reliably different items of the same type (as recognizing different paintings).

As you add more classes an error of only 1-2% might shift another class to the top. If you output the dictionary of results you can see each class and its confidence level -- if all the numbers are relatively close it's having a hard time telling the different classes apart. You may need to give it more examples of each class if that's the case.

I can understand that may fail recognizing 2 different paintings but not so much when it recognizes a plate as a painting or Donald Duck as a seaside. Suppose I want to recognize a painting in a set of 100 different paintings... is this feasible or not?

I think you are giving it too much credit. Many factors can affect the performance, including lighting, angle of incidence relative to the camera, reflective nature of any surfaces involved, etc. Often these neural networks can learn very arbitrary patterns that make no sense to us given our a priori knowledge but are perfectly reasonable w.r.t. the data you've given the model.

I am not giving credit I am testing it. I used the model without changing any parameter and I may suppose that modifing them it could work better but I need to learn more.