WhoFi and Wi-Fi Sensing: Five Questions Behind a Camera-Free AI Claim

Read the experiment before buying the promise

Techweez covered Who-Fi on 29 July 2025. That date belongs to the source report; it is not an ITBank publication date. The underlying research examines person re-identification through Wi-Fi signal measurements rather than camera images. ITBank has not reproduced the experiment or evaluated a commercial deployment.

The primary paper describes a transformer-based approach and tests on the NTU-Fi benchmark, including data from 14 participants. Its reported 95.5% Rank-1 result is a benchmark retrieval measure. It does not demonstrate that an arbitrary router can identify any visitor in an uncontrolled building.

For African offices, schools and other prospective technology buyers, the useful takeaway is an evidence checklist, not a claim that this system is already operating locally.

1. What decision would sensing support?

Begin with the business problem. Counting occupancy, recognising a returning person and identifying a named individual are different requirements. A proposal should explain which is necessary and why a less identifying method would not suffice.

Ask the supplier to state the intended action after a detection. A dashboard that nobody uses can still create data-handling obligations and recurring costs.

2. What was actually measured?

Request the evaluation protocol, participant count, test environment and error definitions. Ask whether test data include people or conditions absent from training. A percentage without its denominator and test design is a weak basis for procurement.

Define acceptable failure in the proposed workflow before accepting a demonstration. An office convenience feature and a decision affecting someone’s access need different scrutiny.

3. What changes outside the demonstration?

Request a bounded, authorised evaluation in representative conditions. State in advance how the team will examine failed detections, mistaken matches and changes in the environment. Keep a manual fallback.

These are proposed evaluation steps, not results claimed by the paper. A successful benchmark is a reason to investigate further, not evidence that maintenance and support disappear.

4. Who can link a signal record to a person?

Removing a camera changes the sensing mechanism. It does not by itself establish anonymity. Ask what records are retained, who can access them, how long they persist and whether they can be linked to other identifiers.

Obtain appropriate local privacy and legal review before any trial involving people. Do not assume that a technology sits outside applicable rules because the identifier is unfamiliar. This article makes no legal determination about a particular deployment.

5. What would make the project stop?

Set a review date and a clear exit condition: inadequate evidence, unacceptable errors, unclear ownership or a less intrusive alternative. Require export and deletion procedures to be demonstrated rather than merely listed in a sales presentation.

The practical lesson is to make claims smaller and tests more precise. Camera-free sensing may warrant research, but useful deployment requires an accountable decision-maker, credible evidence and a clear explanation to the people affected.