Class-Incremental Learning for Semantic Segmentation Re-Using Neither Old Data Nor Old Labels

While neural networks trained for semantic segmentation are essential for\nperception in autonomous driving, most current algorithms assume a fixed number\nof classes, presenting a major limitation when developing new autonomous\ndriving systems with the need of additional classes. In this paper we present a\ntechnique implementing class-incremental learning for semantic segmentation\nwithout using the labeled data the model was initially trained on. Previous\napproaches still either rely on labels for both old and new classes, or fail to\nproperly distinguish between them. We show how to overcome these problems with\na novel class-incremental learning technique, which nonetheless requires labels\nonly for the new classes. Specifically, (i) we introduce a new loss function\nthat neither relies on old data nor on old labels, (ii) we show how new classes\ncan be integrated in a modular fashion into pretrained semantic segmentation\nmodels, and finally (iii) we re-implement previous approaches in a unified\nsetting to compare them to ours. We evaluate our method on the Cityscapes\ndataset, where we exceed the mIoU performance of all baselines by 3.5% absolute\nreaching a result, which is only 2.2% absolute below the upper performance\nlimit of single-stage training, relying on all data and labels simultaneously.\n