High Recall and Weak Localization in Spatially Weighted Camera Soiling Segmentation

Abstract

A vehicle camera soiling detector can achieve high recall by labelling most of an image as contaminated. This study re-examines recorded pixel counts from a U-Net baseline and two spatially weighted training runs, using a fixed 3 × 3 grid and a Gaussian priority map. Each evaluation contains 2,457,600 pixels at a decision threshold of 0.5. Recalculation shows that the Gaussian run detects 99.989% of labelled defect pixels but correctly retains only 0.055% of clean pixels. Its predicted defect area covers 99.972% of the evaluated pixels. Its Dice score of 0.761949 exceeds an all-defect predictor by only 0.00497 percentage points. The baseline Dice score, 0.762570, is slightly below its own all-defect reference; the grid run gives 0.675116. Balanced accuracy and Matthews correlation confirm weak discrimination in all three records. The baseline and weighted evaluations have different foreground totals, so these results are not treated as a controlled comparison of loss functions. An analytical examination also shows that unnormalized spatial weights change the balance between binary cross-entropy and Dice terms. The contribution is a retrospective failure analysis: it identifies extensive oversegmentation, separates this observation from unproven explanations, and specifies the controls needed before spatial weighting can support decisions about vehicle camera availability.