Reproduce “Few-Shot Incremental 3D Object Detection in Dynamic Indoor Environments” 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

Yun Zhu, Jianjun Qian, Jian Yang, Jin Xie, Na Zhao

Researchers introduce FI3Det, the first framework for few-shot incremental 3D object detection, enabling models to adapt to novel object categories with limited samples while preserving performance on previously learned classes. This approach leverages VLM-guided unknown object learning and gated multimodal prototype imprinting to achieve superior detection of novel objects and strong retention of base class knowledge.

Key insights

  • Unannotated 'unknown objects' present in base training scenes, when processed with VLM-guided mining and adaptive weighting, provide valuable auxiliary supervision for improved generalization to novel categories.
  • An adaptive gating mechanism for fusing modality-specific 2D semantic and 3D geometric prototypes enhances robustness and generalization for few-shot novel object detection by balancing cross-modal contributions.
  • Momentum-based prototype imprinting effectively addresses instability and overfitting risks associated with limited novel samples, helping to preserve historical information while learning new concepts incrementally.

Summary by alphaXiv.

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