Both the ftc-ml manual and the email sent out at the beginning of the beta state that we can add additional labels to the TensorFlow models provided in the SDK to recognize objects from the COCO dataset. However, this does not appear to work. I tried adding additional labels but the game objects were still the only objects recognized. Looking into it further, I got the .tflite files from the maven repo and used this .tflite viewer to look at the layers of the model. The final layer appears to restrict the final output to 4 possibilities for the full model or 2 possibilities for the 2 object models, meaning the model will never output something that would correspond to a label outside the normal game objects. It’s possible that I’m doing something the wrong way, but as best I can tell the manual and email are wrong.
I also had an issue with this - I tried adding “CAT” to the LABELS to see if it would work, and then it thought everything in view was a cat. I must be doing something wrong but I was wondering if you ever got this to work? Thanks so much!