Setting the instance detection confidence score threshold

[jkbhagatio]

When running inference, there must somewhere be a confidence score threshold for an instance to be detected right? Let's call it "instance-detection-confidence-score-threshold". Could someone point me to where this is and how this could be adjusted prior to model export?

[roomrys]

Hi @jkbhagatio,

Yes! Sort-of... when loading the model, there is an argument named peak_threshold that allows setting a confidence threshold. However, this threshold is per point not per the entire instance. To read more on the API use of load_model, check out #1603. You can also set the peak_threshold via the sleap-track CLI:

  --peak_threshold PEAK_THRESHOLD
                        Minimum confidence map value to consider a peak as valid.

Furthermore, the peak thresholding is a parameter of inference not training, so it's more like a filter on the output of the model (not integral to the model itself).

Let us know if this helps!

Thanks,
Liezl

[talmo]

Just to add: the peak_threshold parameter should also affect instance detection if you're using top-down models (specifically, it applies to the centroid peak finding).

[jkbhagatio]

Thanks both! Is there a way to set this upon model export, e.g. when using sleap-export cli? I know now there is --max_instances, but still wondering if peak_threshold could be changed

[roomrys]

Hi @jkbhagatio,

There is no way to set peak_threshold when using the sleap-export. As mentioned above:

peak thresholding is a parameter of inference not training, so it's more like a filter on the output of the model (not integral to the model itself)

hence why peak_threshold is only used as an input during inference (and not for sleap-export).

Thanks,
Liezl

[jkbhagatio]

Thanks @roomrys ! Is there a way to do something like --max_instances for id models?