Reproduce “Reinforcement Learning Based Multi-modal Feature Fusion Network for Novel Class Discovery” 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
Qiang Li, Qiuyang Ma, Weizhi Nie, Anan Liu
This paper introduces the Reinforcement Learning Based Multi-modal Feature Fusion Network (RMFFN) for novel class discovery in complex 3D datasets. The framework integrates reinforcement learning, a multi-agent architecture for multi-modal fusion, and a human-inspired pseudo-labeling strategy, yielding superior performance on 3D benchmarks and demonstrating robustness to missing data.
Key insights
- Mimicking human cognitive processes through reinforcement learning allows for more robust and adaptive novel class discovery in unlabeled data.
- Fusing features from multiple 3D modalities (multi-view images, point clouds, meshes, voxels) via a Member-to-Leader Multi-Agent architecture enhances feature understanding and classification.
- A Strict-to-Loose Clustering (STLClu) strategy, which iteratively relaxes pseudo-labeling constraints, effectively generates reliable labels for unknown classes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.