Zero-Fi: Zero-Shot Wi-Fi Activity Recognition via Signal-Language Alignment
A team of researchers has introduced Zero-Fi, a framework designed for contrastive signal-language alignment aimed at recognizing human activities via Wi-Fi without prior training. In contrast to traditional approaches that depend on a fixed set of activities and necessitate labeled Wi-Fi data for each class, Zero-Fi derives unified representations from various Wi-Fi signal features and aligns them with natural-language descriptions of activities within a common embedding space. This innovative alignment allows for the identification of new activity classes without the need for labeled Wi-Fi data or model adjustments. Experiments conducted on extensive public benchmark datasets showcase successful zero-shot recognition of previously unseen activity classes, emphasizing the capability of signal-language alignment to broaden Wi-Fi sensing applications. The findings are available on arXiv under the identifier 2607.26381.
Key facts
- Zero-Fi is a contrastive signal-language alignment framework for zero-shot Wi-Fi-based human activity recognition.
- It learns unified representations from complementary Wi-Fi signal features.
- Aligns them with semantic representations of natural-language activity descriptions in a shared embedding space.
- Enables recognition of new activity classes without labeled Wi-Fi samples or model adaptation.
- Experiments on large-scale public benchmark datasets show effective zero-shot recognition of held-out activity classes.
- Published on arXiv with identifier 2607.26381.
Entities
—