Reproduce “Class-Incremental Few-Shot Object Detection” 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
Pengyang Li, Yanan Li, Han Cui, Donghui Wang
The research introduces LEAST, a method for Class-Incremental Few-Shot Object Detection (iFSD) that allows models to learn new object categories from limited examples while retaining knowledge of previously learned classes. LEAST achieves state-of-the-art performance by balancing detection accuracy across both base and novel classes with efficient resource utilization.
Key insights
- Decoupling the class-agnostic image feature extractor from the class-sensitive object feature extractor effectively balances knowledge retention for base classes and efficient adaptation to novel classes.
- Resource-efficient knowledge distillation via pre-computed logits successfully mitigates catastrophic forgetting without the computational overhead of maintaining a full, active frozen model during training.
- Clustering-based exemplar selection is crucial for preserving the multi-modal distributions of base classes, leading to more robust continual learning performance and reduced degradation over time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.