-
Notifications
You must be signed in to change notification settings - Fork 805
Description
The Problem
The ana-cont models 1300 and newer, 1400 and 1500 have quite bad detection rates especially in the region of zero which is crucial for transition determination and post-processing. Almost every zero crossing is not initially successful and not smooth at all.
- 1300 is quite OK, but still not usable in my opinion
- 1400 is the worst of the three, it detects quite often the complete wrong number
- 1500 is almost as weak as 1400, the variance seems quite high (see example)
I've tested with two different demo sets. Using the same demo sets with ana-cont-1208 (and earlier) or the ana-class100 models, no issues at all.
The last good ana-cont model seems to be 1208 which performs quite well. Still the reference in terms of detection rate is the ana-class100-171. It's little slower than ana-cont but if detection matters for my environment it's still the best.
BTW: Is ana-class100 model not maintained anymore? I've realized that ana-class100 models are not even part of newer releases anymore.
I assume ana-cont models are used quite often, because they are the detault ones, therefore I decided to open this as a bug report. Maybe some others have also trouble with these models and this way it's more visible. Feel free to convert to a discussion.
Version
n.a.
Logfile
n.a.Expected Behavior
No response
Screenshots
Just one example, tested with 1500:
0.7 -> 9.5
9.6 -> 8.6
The variance seems quite high.
Additional Context
No response
