Researchers Misrepresenting the Capability of Human Pose Estimation from WiFi Channel Strength Indicators

Numerous papers in recent years have focused on the estimation of human pose from Wi-Fi channel strength indicators (CSI), with claims of results on par with video techniques. I was able to scrutinize the code from one of the earliest papers and found that the authors had divided the time series data at random into training and test sets. Notably, this practice was not disclosed in their methodology. It’s well understood in Machine Learning that such a practice should be avoided since it leads to data duplication across both sets and results in model overfitting, yielding overly optimistic results.

Surprisingly, these random splits are often omitted from the methodologies of the papers. This raises questions about whether such an oversight is deliberate. Sadly, the measure of an academic’s success is often based on the volume of papers they publish in reputable journals, which could potentially promote lax practices, cheating, and misrepresentation of methodologies to inflate the count of papers authored. The growing number of papers employing this flawed approach makes me suspect that this is not a mere oversight, but a calculated strategy to outperform previous studies. The first paper that adopted this approach set a regrettable precedent, thereby encouraging misconduct in pursuit of publication.

To verify this, one would need to examine the code used to see if these random splits were indeed implemented…