Our new X account is live! Follow @wizwand_team for updates
WorkDL logo mark

Continual Semantic Segmentation via Repulsion-Attraction of Sparse and Disentangled Latent Representations

About

Deep neural networks suffer from the major limitation of catastrophic forgetting old tasks when learning new ones. In this paper we focus on class incremental continual learning in semantic segmentation, where new categories are made available over time while previous training data is not retained. The proposed continual learning scheme shapes the latent space to reduce forgetting whilst improving the recognition of novel classes. Our framework is driven by three novel components which we also combine on top of existing techniques effortlessly. First, prototypes matching enforces latent space consistency on old classes, constraining the encoder to produce similar latent representation for previously seen classes in the subsequent steps. Second, features sparsification allows to make room in the latent space to accommodate novel classes. Finally, contrastive learning is employed to cluster features according to their semantics while tearing apart those of different classes. Extensive evaluation on the Pascal VOC2012 and ADE20K datasets demonstrates the effectiveness of our approach, significantly outperforming state-of-the-art methods.

Umberto Michieli, Pietro Zanuttigh• 2021

Related benchmarks

TaskDatasetResultRank
Semantic segmentationPascal VOC 15-1 setting 2012 (val)
mIoU (all)48.7
88
Semantic segmentationPascal VOC 15-5 setting 2012 (val)
mIoU (All)70.1
82
Semantic segmentationPascal VOC overlapped setting (15-1 (6 steps))
mIoU (Classes 1-15)44.7
41
Continual Semantic SegmentationPascal-VOC 15-1 scenario 2012
mIoU (classes 0-15)47.3
32
Continual Semantic SegmentationPascal-VOC 15-5 scenario 2012
mIoU (Classes 0-15)76.3
30
Incremental Semantic SegmentationPascal VOC disjoint setup 2012 (VOC 10-1)
mIoU (0-10)67.5
30
Semantic segmentationPascal VOC 19-1 setting 2012 (val)
mIoU (old)71.3
29
Continual Semantic SegmentationADE20k 100-50 (2 tasks) (val)
mIoU (0-100)40.52
28
Continual Semantic SegmentationPascal-VOC 19-1 2012
mIoU (0-19 Classes)71.3
27
Semantic segmentationADE20k overlapped setting (100-10 (6 steps))
mIoU (Classes 1-100)29.9
21
Showing 10 of 29 rows

Other info

Code

Follow for update