Generation and evaluation of underwater out-of-distribution sonar signals using generalized virtual outlier synthesis
We introduce GVOS, a method that uses Gaussian mixture model (GMM) to generate out-of-distribution (OOD) data for passive sonar signals more accurately than existing approaches.
GVOS generates virtual outliers by sampling from the low-likelihood regions of the estimated Gaussian mixture distribution in the feature space.
GVOS accurately estimates the distribution and captures data characteristics in low-likelihood regions, resulting in the generation of more precise virtual outliers in experiments.
