End-to-end Solution for Tenebrio Molitor Rearing Monitoring with Uncertainty Estimation and Domain Shift Detection

Paweł Majewski, Piotr Lampa, Robert Burduk, Jacek Reiner; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 5498-5507

The large-scale rearing of edible insects of which Tenebrio Molitor is a representative requires monitoring using vision systems to control the process and to detect anomalies. Previously proposed solutions by researchers relied on multiple modules related to specific tasks (calculated coefficients) and specific types of models (instance segmentation semantic segmentation). Long processing times and difficulties in maintaining and updating modules encourage the search for a more condensed solution as an end-to-end model. This paper proposed a modified YOLOv8 architecture extended with additional heads related to specific tasks. Heads were trained on problem-oriented small datasets which significantly reduced the time spent on sample annotation. The proposed solution also included estimation of prediction uncertainty based on variation among predictions in model ensemble and detection of domain shift phenomenon. Quantitative results from the conducted experiments confirmed the potential of the developed solution.