Parametric and Multivariate Uncertainty Calibration for Regression and Object Detection

Reliable spatial uncertainty evaluation of object detection models is special interest and has been subject recent work. In this work, we review the existing definitions for calibration probabilistic regression tasks. We inspect properties common networks extend state-of-the-art recalibration methods. Our methods use a Gaussian process (GP) scheme that yields parametric distributions as output (e.g. or Cauchy). The usage GP allows local (conditional) by capturing dependencies between neighboring samples. such simplified adaption in subsequent processes, e.g., Kalman filtering scope tracking. addition, to perform covariance estimation which post-hoc introduction correlations quantities, position, width, height detection. To measure joint multivariate possibly correlated data, introduce quantile error based on Mahalanobis distance predicted distribution ground truth determine whether within quantile. experiments show overestimate comparison observed error. simple Isotonic Regression method sufficient achieve good quantification terms calibrated quantiles. contrast, if normal are required our GP-Normal best results. Finally, able results calibration. All code open source available at https://github.com/EFS-OpenSource/calibration-framework .