Then boundaries_id = 'gt' the boundaries are converted first to frame indices here. Then once the labeling algorithm outputs the results this line recomputes the boundaries from the frame indices, potentially adding an end-of-audio boundary.
If however the end-of-audio boundary is present in the ground truth boundaries, it becomes doubled after this line is executed. This is due to some discrepancy in the conversion to frame indices and back to time.
Here is an example
est_times of ground truth input boundaries: [0, 12.689583333333333, 27.68958333333333, 42.18958333333333, 56.380208333333336, 70.63020833333333, 83.38020833333333, 98.88020833333333, 125.88020833333333, 145.33020833333333, 157]
est_indices:
[ 0 274 597 909 1215 1521 1796 2130 2711 3130 3381]
est_times:
[ 0. 12.72453515 27.72462585 42.21387755 56.4244898
70.63510204 83.4060771 98.9170068 125.8985941 145.3569161
157.01333333 157.058322 ]
here the config:
<class 'dict'>: {'annot_beats': False, 'feature': 'pcp', 'framesync': True, 'dirichlet': False, 'xmeans': False, 'k': 4, '2dfmc_offset': 4, 'label_norm_feats': inf, 'label_norm_floor': 0.1, 'label_norm_min_db': -80, 'features': <msaf.features.PCP object at 0x124eba8d0>, 'hier': False}
Then boundaries_id = 'gt' the boundaries are converted first to frame indices here. Then once the labeling algorithm outputs the results this line recomputes the boundaries from the frame indices, potentially adding an end-of-audio boundary.
If however the end-of-audio boundary is present in the ground truth boundaries, it becomes doubled after this line is executed. This is due to some discrepancy in the conversion to frame indices and back to time.
Here is an example
here the config: