Smooth trajectories using a median filter

[niksirbi]

DeepLabCut has some support for smoothing/filtering pose tracks, in the form of the filterpredictions() function in the filtering.py module.

They offer three options:

• Median filtering. This is the most straightforward and the easiest to understand + implement. It relies on the scipy.signal.medfilt().
• Spline filtering. This method will identify gaps in the tracks and interpolate them using splines. Will be a bit harder to adapt the code for your SLEAP tracks, but can be done.
• Arima filtering. Under the hood this uses the SARIMAX model from the statsmodels package. This will be more complicated, and I honestly don’t yet understand how it works.

We should also offer several filtering options eventually, but the median filter is probably the easiest to start with.

Maybe a first step would be to create a filtering.py/smoothing.py module and add a single median filter function. That function should:

• take the pose tracks xarray.Dataset as input
• apply a median filter along the time axis (using either scipy.signal.medfilt() or scipy.ndimage.median_filter())
• return a filtered pose tracks xarray.Dataset as output
• allow to specify the window length (with sensible default value)
• perhaps utilize the confidence scores (if present) in some clever way
• be efficiently implemented in a way that's xarray-aware (use xarray's built-in functionalities as much as possible)
• may be chunked-parallelized via Dask (though this may be also left for a future version)

[niksirbi]

Primer on mean/median filtering for timeseries data.

[niksirbi]

In xarray, this could be as simple as doing:

da.rolling(time=5).median()

See rolling-window-operations.

If we want to be fancy, we can set the timepoints below a specified confidence threshold to NaN, before applying the median filter.