Subclassified Loss: Rethinking Data Imbalance From Subclass Perspective for Semantic Segmentation

Abstract

Semantic segmentation plays a crucial role in enabling intelligent vehicles to perceive and understand their surroundings. However, datasets used for semantic segmentation often suffer from data imbalance, where the number of pixels belonging to different classes varies significantly. To address this challenge, various novel loss functions have been proposed at the class or pixel level to counterbalance the data imbalance. In this study, we propose a novel approach to mitigate this problem from a subclass perspective. Specifically, we identify subclasses within each class based on the similarity of feature maps. These subclasses are then reweighted according to their distribution. Our approach can seamlessly transition into widely used loss functions, such as cross-entropy loss and class-weighted cross-entropy loss, in extreme cases. The proposed loss function is compatible with existing semantic segmentation methods, serving as a plug-in component for both current and future methodologies. We conducted extensive experiments on challenging datasets, namely SemanticKITTI and Cityscapes, to evaluate the effectiveness and generalization of our subclassified loss. Our experiments involved LiDAR-based and image-based methods, including RangeNet++, KPRNet, PointRend, STDC, and SegFormer, resulting in substantial improvements in terms of intersection-over-union (IoU) for segmentation. <supplementary-material orientation="portrait" position="float" specific-use="print" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> The codes are publicly available at <uri>https://github.com/skyshoumeng/Subclassified-Loss</uri> </supplementary-material>

Topics & Concepts