I am part of a team on developing CNNs to detect species in the Peruvian Amazon from soundscapes. We are having trouble finding a ground truth dataset for the region to test our model against. We have been trying to use “A collection of fully-annotated soundscape recordings from the Southwestern Amazon Basin” (https://zenodo.org/record/7079124#.Y7iis-xudhE) which seems to be the best publicly available strongly labeled ground truth from the region. However, we think there are a few errors in the labeling leading to our models having poorer results than they might. For instance, in the screenshot below, there seems to be an unlabeled species between 34:38 and 34:30 from 1 to 4 hz (and repeated again at 34:35 to 34:38) that is labeled in annotations 15616 and 15622. We have noticed a handful of what we think are these errors in the dataset, but we don’t necessarily have the expertise to know for sure. Has anyone used this data in training/testing and if so what are your thoughts on this data?
Thank you for your time!