Reproduce “Adaptive Visual Scene Understanding: Incremental Scene Graph Generation” 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
Naitik Khandelwal, Xiao Liu, Mengmi Zhang
Researchers at NTU and A*STAR introduced Continual Scene Graph Generation (CSEGG) as a new problem, developing a comprehensive benchmark and proposing "Replays via Analysis by Synthesis" (RAS). This method, which synthesizes replay images from symbolic scene graphs, consistently outperforms existing continual learning baselines in adapting scene graph models to new objects and relationships, demonstrating strong generalization capabilities while reducing memory footprint.
Key insights
- Continual learning for Scene Graph Generation (SGG) faces unique challenges, including severe catastrophic forgetting, which are not adequately addressed by conventional continual learning methods.
- Generative replay leveraging symbolic scene graph representations can synthesize diverse and semantically coherent replay data, outperforming raw image replay for complex tasks like SGG.
- Disentangling relational knowledge from specific object identities through symbolic generation promotes better generalization of known relationships to previously unseen objects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.