Realistic Predictors for Regression and Semantic Segmentation

Abstract

Computer vision and image processing algorithms work well under strong assumptions. Computer vision algorithms are not expected to do well on all kinds of inputs. For instance, excessively noisy images may not fetch optimal results for most computer vision algorithms. Unexpected outputs from the computer vision module can have negative downstream consequences for other modules in the pipeline. To mitigate such consequences, we use a predictor framework that was simultaneously trained with a Hardness Predictor network. This framework guarantees improved performance over those images with lower "hardness" values. The proposed predictor framework, when applied to the input data, would result in a relatively lower variance estimator when the size of the training set is large, both in the domain of semantic segmentation as well as regression analysis.