Inherit With Distillation and Evolve With Contrast: Exploring Class Incremental Semantic Segmentation Without...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Class Incremental Semantic Segmentation (CISS) faces challenges with catastrophic forgetting and semantic drift.
• Existing knowledge distillation methods struggle with pixel confusion due to missing annotations.
• Data-replay methods incur storage and privacy issues.

Purpose of the Study:

• To address CISS without relying on exemplar memory.
• To synchronously resolve catastrophic forgetting and semantic drift.
• To improve semantic segmentation performance in incremental learning scenarios.

Main Methods:

• Proposed Inherit with Distillation and Evolve with Contrast (IDEC) framework.
• Introduced Dense Knowledge Distillation on all Aspects (DADA) for semantic-invariant knowledge inheritance.
• Implemented Asymmetric Region-wise Contrastive Learning (ARCL) to mitigate semantic drift.

Main Results:

• Achieved state-of-the-art performance on Pascal VOC 2012, ADE20K, and ISPRS datasets.
• Demonstrated superior anti-forgetting capabilities, especially in multi-step CISS tasks.
• Effectively resolved catastrophic forgetting and semantic drift without exemplar memory.

Conclusions:

• The IDEC method offers an effective solution for CISS without data replay.
• DADA and ARCL modules successfully preserve knowledge and reduce semantic drift.
• The proposed approach significantly advances the field of incremental learning for semantic segmentation.