Incremental Learning Techniques for Semantic Segmentation

31 Jul 2019  ·  Umberto Michieli, Pietro Zanuttigh ·

Deep learning architectures exhibit a critical drop of performance due to catastrophic forgetting when they are required to incrementally learn new tasks. Contemporary incremental learning frameworks focus on image classification and object detection while in this work we formally introduce the incremental learning problem for semantic segmentation in which a pixel-wise labeling is considered. To tackle this task we propose to distill the knowledge of the previous model to retain the information about previously learned classes, whilst updating the current model to learn the new ones. We propose various approaches working both on the output logits and on intermediate features. In opposition to some recent frameworks, we do not store any image from previously learned classes and only the last model is needed to preserve high accuracy on these classes. The experimental evaluation on the Pascal VOC2012 dataset shows the effectiveness of the proposed approaches.

PDF Abstract

Results from the Paper

Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Overlapped 100-5 ADE20K ILT mIoU 0.5 # 7
Domain 11-5 Cityscapes ILT mIoU 59.1 # 5
Domain 1-1 Cityscapes ILT mIoU 30.1 # 5
Domain 11-1 Cityscapes ILT mIoU 57.8 # 4
Disjoint 15-1 PASCAL VOC 2012 ILT mIoU 7.9 # 7
Overlapped 15-5 PASCAL VOC 2012 ILT Mean IoU (val) 61.3 # 11
Overlapped 15-1 PASCAL VOC 2012 ILT mIoU 9.2 # 11
Disjoint 15-5 PASCAL VOC 2012 ILT Mean IoU 58.9 # 7
Overlapped 10-1 PASCAL VOC 2012 ILT mIoU 5.5 # 11
Disjoint 10-1 PASCAL VOC 2012 ILT mIoU 5.4 # 7


No methods listed for this paper. Add relevant methods here