Determine recall for large-scale deployment of the multiclass ResNet50 detector

Determine the recall (true positive rate) of the deployed multiclass ResNet50 convolutional neural network detector for female gibbon calls when applied to the wide passive acoustic monitoring arrays at Danum Valley Conservation Area (Malaysia) and Jahoo (Cambodia) under real-world operating conditions and a 0.9 confidence threshold, in order to quantify detection sensitivity at scale.

Background

The paper deployed a multiclass ResNet50 model trained with duplicated data and color jitter over extensive passive acoustic monitoring arrays in Danum Valley and Jahoo, using a 0.9 confidence threshold. The authors manually labeled detections to estimate precision, reporting approximately 0.90 precision in Danum Valley and 0.89 in Jahoo.

However, because of the large search space and practical constraints in fully annotating all ground truth events across the wide arrays, the authors were unable to effectively estimate recall for these large-scale deployments, leaving the detector’s sensitivity under operational conditions unresolved.