search

UCSD-E4E / PyHa

A repo designed to convert audio-based "weak" labels to "strong" intraclip labels. Provides a pipeline to compare automated moment-to-moment labels to human labels. Methods range from DSP based foreground-background separation, cross-correlation based template matching, as well as bird presence sound event detection deep learning models!
Other
16 stars 10 forks source link

Integrate BirdNET Foreground-background separation #174

Closed JacobGlennAyers closed 5 months ago

JacobGlennAyers commented 9 months ago

Overview - image

Taken from this source: https://ceur-ws.org/Vol-1609/16090547.pdf

JacobGlennAyers commented 9 months ago

I would want the isolation parameters to look something like this -

isolation_parameters = { "model" : "power_fg-bg_separation", "technique" : "simple", "threshold_type" : "pure", "threshold_const" : 0.5, "verbose" : True } These exact parameters could reproduce their work. With that being said, we could consider adding in some other hyperparameters. For instance, the size of the erosion and dilation kernels is a hyperparameter. Furthermore, the time and frequency multiplier thresholds to determine an active pixel. Currently any pixel 3x greater than the row median and 3x greater than the column median are set active.

"temporal_threshold" : 3 "freq_threshold" : 3 "kernel_size" : 4

JacobGlennAyers commented 9 months ago

https://github.com/UCSD-E4E/PyHa/pull/175

JacobGlennAyers commented 5 months ago

Recently Merged into main