Spectral Analysis for Semantic Segmentation with Applications on Feature Truncation and Weak Annotation

We propose spectral analysis to investigate the correlation between accuracy and resolution of segmentation maps for semantic segmentation. The current networks predict on down-sampled grid images alleviate computational cost. Moreover, these can be trained by weak annotations that utilize only coarse contour maps. Despite successful achievement works utilizing low-frequency information maps, however, resultant may also degraded in regions near object boundaries. It is yet unclear a theoretical guideline determine an optimal strike balance cost analyze objective function (cross-entropy) network back-propagation process frequency domain. discover cross-entropy key features CNN are mainly contributed components This further provides us quantitative results efficacy then validated two applications: feature truncation method block-wise annotation limit high-frequency annotation, respectively. agree with our analysis. Thus success existing work now has foundation.