Probabilistic Prediction and Uncertainty for Underwater 3-D Reconstruction

Predicting the performance of complex “black-box” algorithms such as underwater 3D reconstruction is a core problem in instrumentation and measurement when these algorithms are deployed as virtual instruments in autonomous underwater vehicle (AUV) perception pipelines. Conventional reactive sensing cannot anticipate how current acquisition decisions will affect future reconstruction quality, limiting the ability to manage risk and onboard resources. We propose a forward-looking probabilistic performance measurement framework that treats a given reconstruction pipeline as a black-box instrument and forecasts the joint distribution of its future accuracy and completeness on upcoming sequence segments. The framework combines a dual-stream spatio-temporal encoder that models appearance and motion, a Prospective Efficacy Accumulator that distills performance-relevant factors, and a multi-horizon probabilistic decoder that produces calibrated performance distributions across prediction horizons. A geometry-aware domain adaptation module further preserves multi-view consistency in sonar-to–pseudo-optical translation, ensuring that the front-end conditioning stage remains compatible with downstream reconstruction. Experiments on synthetic SeaWorld-Sim data and a public real AUV forward-looking sonar dataset demonstrate that the proposed virtual performance sensor consistently improves mean absolute error and expected calibration error over strong deterministic and probabilistic baselines for multiple reconstruction back-ends. The resulting predictive distributions act as decision-ready measurement outputs—expected accuracy/completeness together with quantified uncertainty—that support risk-aware planning and resource allocation in underwater 3D reconstruction missions.