Contrastive learning framework tackles multimodal human activity recognition with limited data

Researchers have developed CLMM, a new contrastive learning framework designed for multimodal human activity recognition, particularly when labeled data is scarce. The framework utilizes a two-stage training process, first capturing shared cross-modal information with a CNN-DiffTransformer encoder and a novel weighting algorithm, then focusing on modality-specific features with a dual-branch architecture. Experiments on public datasets show CLMM surpasses existing methods in both recognition accuracy and convergence speed. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a novel framework for multimodal recognition with limited data, potentially improving applications relying on human activity analysis.

RANK_REASON This is a research paper detailing a new framework for human activity recognition.

Read on arXiv cs.CV →

• CLMM
• CNN
• DiffTransformer
• arXiv

• paper
• other