Contour-Aware Equipotential earning for Semantic Segmentation

With increasing demands for high-quality semantic segmentation in the industry, hard-distinguishing boundaries have posed a significant threat to existing solutions. Inspired by real-life experience, i.e., combining varied observations contributes higher visual recognition confidence, we present equipotential learning (EPL) method. This novel module transfers predicted/ground-truth labels self-defined potential domain learn and infer decision along customized directions. The conversion is implemented via lightweight differentiable anisotropic convolution that neither incurs parameter overhead of neural networks nor changes architectures. Besides, designed two loss functions, point line implement field regression category-level contour learning, respectively, enhancing prediction consistencies inter/intra-class boundary areas. More importantly, EPL agnostic network architectures, thus it can be plugged into most models. paper first attempt address problem with learning. Meaningful performance improvements on Pascal Voc 2012 Cityscapes demonstrate proposed benefit off-the-shelf fully convolutional models when recognizing intensive comparisons analysis show favorable merits distinguishing semantically-similar irregular-shaped categories.