Reproduce “ConSept: Continual Semantic Segmentation via Adapter-based Vision Transformer” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Bowen Dong, Guanglei Yang, Wangmeng Zuo, Lei Zhang

ConSept is a Vision Transformer-based framework for continual semantic segmentation that integrates lightweight attention-based adapters and specific distillation strategies to learn new categories without forgetting old ones. The method achieved state-of-the-art performance on PASCAL VOC and competitive results on ADE20K, demonstrating robust anti-catastrophic forgetting and improved generalization to novel classes.

Key insights

  • Vanilla pre-trained ViTs, even with a simple linear head, struggle significantly with generalizing to novel classes in a continual learning context without specialized adaptation mechanisms.
  • Lightweight attention-based adapters, when integrated into a ViT, can substantially improve novel class generalization (e.g., 26.51% mIoU increase) with minimal additional parameters, even with a frozen backbone.
  • Full fine-tuning of the entire ViT feature extractor, combined with carefully designed distillation (especially on deeper features), is crucial for effectively learning new concepts while mitigating catastrophic forgetting of old knowledge.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.