0.5.0
This is the first stable release of the 0.5 series. The main change with respect to the v0.4 series is a modification of the flight call classification algorithm which enforces hierarchical consistency in the TaxoNet output. We now ensure that the predicted species belongs to the predicted family and that the predicted family belongs to the predicted order. Additionally, BirdVoxDetect now reports order confidence, family confidence, and species confidence in addition to detection confidence. More details below.
Our numerical experiments on 300 hours of annotated audio have demonstrated that TaxoNet outperforms the classifier of BirdVoxDetect v0.3 (see previous release notes for details). However, we have noted that the output of TaxoNet is occasionally inconsistent across levels. For example, the "family" level may return "Turdidae" (Thrushes) while the "species" level would return "WTSP", i.e. White-Throated Sparrow.
Although infrequent, this lack of consistency may be confusing to users. To address it, we have implemented a new algorithm for hierarchical classification which ensures consistency by design: see #73
The key idea is to query TaxoNet in a coarse-to-fine fashion, while progressively narrowing down the list of candidate species corresponding to the detected flight call. Thanks to Justin Salamon for proposing this elegant solution.
We have taken the opportunity of rewriting the classify_species function (see #73) to enrich the information that is stored in the BirdVoxDetect checklist. In addition to textual labels for order, family, and species, we have added three new columns, corresponding to "order confidence", "family confidence" , and "species confidence".
These values of confidence all range between 50% and 100% and provide insight on the prediction of TaxoNet. Users should be warned, however, that the percentages are not calibrated, and thus do not come with any probabilistic interpretation. Rather, they are simply the output values of the sigmoid units in TaxoNet. This is unlike "Detection confidence (%)", which is carefully calibrated to match the precision of the detector (on a precision-recall curve).