CSI BRIDGE / Ongoing
CSI BRIDGE
Can an invisible wireless signal be described in words? CSI BRIDGE explores a bridge between Wi-Fi Channel State Information and vision-language representations. The aim is to describe actions and situations using only CSI as input at inference time.
CSI BRIDGE
Why CSI-to-Caption?
A fixed action label can leave much of a situation's context unexplained. This research explores connecting movement-related changes in wireless signals to natural language descriptions of what is happening. CSI can itself reveal sensitive behavioral information, so consent and the scope of data use also matter.
Research goals
Study the relationship between temporal changes in Wi-Fi CSI and VLM semantic representations, alongside representations and evaluation methods for CSI-to-Caption. Generalization across people and environments, and agreement between descriptions and observed actions, remain key questions.
Problem & core idea
The challenge is to move beyond selecting a fixed activity label and connect wireless signals to descriptions of a situation. The proposed direction transfers semantic representations from a training-time Teacher VLM to a Student CSI model.
Dataset
The dataset, collection scale and training/evaluation splits have not been made public. Alignment of signals with visual and language information, participant consent and permitted data use are being considered together.
Evaluation direction
The intended questions are whether captions match actual behavior and whether descriptions remain useful across people and environments. Metrics and comparison protocols are under consideration; no validated scores or experimental results are published here yet.
MULTIMODAL REPRESENTATION ALIGNMENT
A research design connecting signals and meaning
This is an outline of the approach under investigation, not a claim of a finalized implementation or validated performance.
01 / Wi-Fi CSI
Examine motion-related changes in wireless channel information, including amplitude and phase. Consider how time alignment, noise, and environmental changes affect the input representation.
02 / Teacher VLM
During training, explore representations from a Teacher VLM that interprets corresponding visual information and language. The teacher provides a route to associating signal patterns with meaning.
03 / Representation Alignment
Investigate how CSI and vision-language representations can express shared semantics. Corresponding data and the design of learning objectives are central to this multimodal alignment.
04 / Student CSI Model
A student processes CSI and aims to learn the teacher's semantic representations, towards describing actions and situations without visual input at inference time.
Possible applications
Potential directions include indoor activity understanding, interaction between people and spaces, and assistance with robot situational awareness. Reliability in real environments and suitability for medical or safety use would require separate validation.
What is public today
This page introduces the research topic and intended approach. Datasets, implementation details, performance figures, and comparative results have not yet been published here. Research notes and results will be added as they become ready to share.