AAI_2025_Capstone_Chronicles_Combined

domain-specific calibration strategy ensures that high-confidence predictions can be trusted, while lower-confidence cases are flagged for human verification.

The project’s semi-automated pipeline design—from the modular training infrastructure and rich augmentation strategy to the Gradio-powered human-in-the-loop interface—ensures that WildScan is not only accurate but also scalable, adaptable, and actively improving over time. The confidence-based uncertain sample identification and TvA calibration mechanisms together form a dual-layer quality control system: one that optimizes human effort and the other that guarantees trustworthy probabilistic outputs for automated decision-making. Overall, WildScan represents a field-ready, semi-automated AI pipeline that bridges the gap between research-grade wildlife monitoring and deployable conservation tools. Its combination of robust detection, competitive classification accuracy, intelligent active learning, and strategic TvA calibration integration positions it as a practical solution for biodiversity monitoring that can scale with conservation needs while minimizing human workload. The system’s ability to continuously learn from human feedback on uncertain cases and to maintain calibrated confidence estimates ensures adaptive, reliable performance—making WildScan a valuable platform for sustainable, long-term wildlife analytics and conservation decision support. Böhner, H., Kleiven, E. F., Ims, R. A., & Soininen, E. M. (2023). A semi-automatic workflow to process images from small mammal camera traps. Ecological Informatics, 74 , 102150. https://doi.org/10.1016/j.ecoinf.2023.102150 Dolan, E., Koire, S., & Woods, G. (2020). Automatic species classification in camera trap images for new locations: A computer vision project [Unpublished course project]. Stanford University, CS230. https://cs230.stanford.edu/projects_spring_2020/reports/38838639.pdf References

300