I am training a model which predicts multiple labels for same object. Example if it is a red car, I need two labels red and car. I trained it by giving different class numbers for the same bounding box in labels. While detecting on test images locally, I get to see multiple labels. However when detecting live on my OAK-D Pro, I always see a single label. I was wondering why I am not able to detect multi labels on the camera, whereas I am able to detect locally. Please help me with it.

Thanks.

  • erik replied to this.

    Hi MonalisaAchalla ,
    Multi-labels aren't yet supported for on-device decoding, perhaps you could have multiple labels for all permutations, eg

    1=red car
    2=blue car
    3=yellow car
    4=red truck
    5=blue truck
    ...

    It's not the cleanest solution, but should work. Another approach would be to run classification after the detection part (so classification for the color after detecting vehicle). Another approach would be to decode the multi-label NN results on the host instead of on-device. Thoughts?
    Thanks Erik