Calibration dataset

Hi,

I’ve reviewed the DFC guide but couldn’t find a clear explanation of what the calibration dataset should include. I have a few questions I’m hoping you can help clarify:

  • Should the calibration images come from the training set?
  • What should the calibration set size be?
  • When you mention “representative images,” could you clarify what that means? For instance, my model will be running in a hallway environment all day. While the training data included both general images and some captured in the target location, I want to ensure the calibration set is appropriate.

Thanks in advance for your help!

You can use images from the training, validation or test set.

Please work trough the tutorials in the Hailo AI Software Suite Docker and read the Model Optimization and Accuracy Analysis chapters in the Hailo Dataflow Compiler User Guide.

It means you should get a good mix of images with different objects and lighting conditions … and not x number of images that are almost identical.
If your dataset is has images that are not well distributed e.g. 5000 dogs and only 100 cats you may want to manually included some images that are not so common instead of just picking all images randomly.

1 Like