Generate own models from images

Will be implemented with the SDK the process for generate own models from images and that can do classification/recognition?

3replies Oldest first
  • Oldest first
  • Newest first
  • Active threads
  • Popular
  • Great question! Right now we are offering pre-trained models, but it's on our roadmap for future releases to be able to offer this capability 

    Reply Like
  • It would also be nice to train models for detection and segmentation.

     

    Just ran the SDK on a detection dataset.  It is blazing fast!  But the accuracy/precision/recall/f1 don't look as good as other models that I've trained on my data.  The SDK model is about 5x faster, but the f1 score is 35% lower than the other model I'm looking at.

     

    In the future will users be able to trade-off accuracy with speed?  How fine-grained is that trade-off?  Are users switching out models or pulling a slider?

    Reply Like
    • Chris Hall Hi Chris.

      In the future will users be able to trade-off accuracy with speed? 

      Absolutely. In future releases, there will be a constellation of models where you can bias for accuracy, performance, or perhaps both.

      How fine-grained is that trade-off?

      TBD, but I imagine it to be akin to t-shirt sizes: Small, Medium, Large, XL. Larger models are more accurate.

      But the accuracy/precision/recall/f1 don't look as good as other models that I've trained on my data. 

      Keep in mind that these models that we've trained aren't particularly tuned for a particular use case. We'd definitely see a marked improvement on those metrics given a  use case, e.g., home security.

      Happy to take this offline Chris. Please feel free to e-mail me directly.

      Reply Like
Like1 Follow
  • 1 Likes
  • 4 mths agoLast active
  • 3Replies
  • 82Views
  • 5 Following