Reproduce “Few-Shot Classification of Interactive Activities of Daily Living (InteractADL)” on OpenResearch

About this paper

Zane Durante, Robathan Harries, Edward Vendrow, Zelun Luo, Yuta Kyuragi, Kazuki Kozuka, Li Fei-Fei, Ehsan Adeli

Researchers from Stanford University and Panasonic Corporation introduced InteractADL, a multi-view dataset of complex, interactive Activities of Daily Living, designed to benchmark fine-grained recognition in real home environments. They also developed Name Tuning and CoNa, methods for Visual-Language Models that optimize class name representations, outperforming traditional prompt tuning (CoOp) by 1.6% to 2.2% absolute on various few-shot activity recognition tasks and achieving state-of-the-art results on MOMA-LRG.

Key insights

• Directly optimizing class name representations in Visual-Language Models (VLMs) through Name Tuning provides greater semantic flexibility and separability, leading to improved few-shot classification performance compared to methods that only tune context prompts.
• InteractADL proves to be a highly challenging benchmark for state-of-the-art VLM-based methods, with overall accuracies ranging from 3% to 7%, indicating substantial opportunities for future research in fine-grained interactive ADL recognition.
• The proposed methods, Name Tuning and CoNa, effectively generalize across various VLMs and datasets (including image classification), often outperforming baselines without requiring extensive video pre-training or large meta-training sets.

Summary by alphaXiv.