Layer-Specific Knowledge Distillation for Class Incremental Semantic Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Class incremental semantic segmentation (CISS) faces catastrophic forgetting.
• Knowledge distillation (KD) is used to mitigate forgetting but lacks layer-specific adaptation.
• Existing KD methods use uniform distillation schemes and fixed weights, ignoring feature characteristics.

Purpose of the Study:

• To propose a layer-specific knowledge distillation (LSKD) method for CISS.
• To enhance KD effectiveness by considering feature characteristics of different intermediate layers.
• To improve performance in open-world semantic segmentation settings.

Main Methods:

• Developed a layer-specific knowledge distillation (LSKD) approach.
• Introduced mask-guided distillation (MD) to handle background shifts.
• Implemented mask-guided context distillation (MCD) for high-level semantic features.
• Utilized a regularized gradient equilibrium method for dynamic trade-off weights.
• Employed bi-level optimization to learn distillation schemes and weights simultaneously.

Main Results:

• LSKD demonstrated superior performance compared to existing methods.
• Achieved state-of-the-art results on Pascal VOC 12 and ADE20K datasets.
• Effectively alleviated catastrophic forgetting in CISS.

Conclusions:

• LSKD offers a more effective approach to knowledge distillation for CISS.
• Layer-specific distillation strategies significantly improve segmentation performance.
• The proposed methods address limitations of current KD techniques in incremental learning scenarios.