Author(s): Alexandra Kolosova; Jeffrey A. Tuhtan

Linked Author(s): Jeffrey Tuhtan

Keywords: Conformal prediction; Fish detection; Deep learning; Computer vision

Abstract: The ecohydraulics community is increasingly making use of real-time machine learning systems for fish video monitoring because these systems can substantially reduce video data volume and decrease the need to store or transmit videos to a central server for further processing. Despite this major advance, models still rely on human verification and retraining, both of which require human raters to perform time-intensive manual annotations. On average, human raters annotate approximately 500 bounding boxes per hour, with experienced annotators reaching up to 1,000 boxes per hour. The novelty of this work is the application of a conformal prediction framework to provide statistically rigorous uncertainty quantification of model-predicted bounding boxes containing fish, allowing for a substantial reduction in the manual annotation effort. In addition, conformal prediction replaces the commonplace heuristic methods used to assess model-derived confidence and guarantees that valid bounding boxes with fish are included in the prediction sets at a user-defined error level. To test our method, we evaluated it at significance levels of practical importance (99%, 95%, 90%, and 80%) and analyzed its coverage and prediction set size. We discuss limitations related to model calibration, nonconformity measure selection, and prediction set precision . Using our approach, the required time for manual annotation was reduced by up to a factor of 28.5. This work is significant because it provides a fully automated method to rigorously identify true fish detections in large datasets with high statistical confidence, substantially reducing the number of detections that human raters need to review for verification, and can aid in selecting data for retraining.

DOI: https://doi.org/10.64697/iahr.proc.ise2026.abs.186

Year: 2026