AI measurement and prediction of the suitable firmness level

A two-dimensional photo shows a projection. Perspective, clothing, lighting, and body posture influence the dimensions estimated from it. A three-dimensional scan can provide more shape information but can also have measurement errors. Neither a photo nor a surface scan automatically captures body mass or the mechanical properties of deep tissue.

The first quality question is, therefore, which features are actually measured and which are only estimated. Shoulder width, waist circumference, and visible contour are not the same as the load in the lateral position. An algorithm should not hide missing information behind an overly precise firmness specification.

Recognizing body posture, estimating pressure distribution, and predicting the most comfortable mattress are different tasks. The research paper BodyMAP investigates the joint estimation of a body model and three-dimensional pressure distribution for people in bed. [1] It shows technical possibilities but is not general proof of an ideal sales decision based on a normal customer photo.

Methods for sleep position recognition based on few pressure sensors also primarily address a classification task. [2] High accuracy for supine or lateral position says nothing about whether the recommended mattress reduces pain in the long term. A product must be evaluated according to exactly the goal it promises later.

If the model learns from previous sales, it can primarily reproduce previous sales practices. If it learns from pressure minima, it may optimize pressure at the expense of mobility or support. If it predicts satisfaction, brand, price, and service may influence the target variable.

A meaningful dataset should therefore contain several separate endpoints: comfort, position, movement, complaints, and, if applicable, objective measurements. Furthermore, the offered mattress range must be known. A firmness level H3 from a manufacturer is not a universal mechanical standard. The recommendation should be related to specific models or measured properties.

A model must be tested on new individuals who do not appear in similar form in the training. If photos of the same person are distributed across training and testing, performance assessment can be too optimistic. New cameras, different clothing, or a changed product range can also reduce accuracy.

For STOLL, results based on body build, age, mobility, and relevant application situations would be important. High average accuracy can mask poor results for smaller groups. The system should be able to reject a recommendation or request additional measurements if the data lies outside its validated range.

A useful explanation names the decisive features and the remaining uncertainty. "The shoulder zone might need more compliance" is more helpful for a lying test than an opaque overall rating. The final selection takes the person's feedback into account and remains correctable.

Body images are personal data and may contain particularly sensitive information depending on processing. Not every photo is automatically biometric data in the special legal sense; however, an identification or health analysis can change the requirements. Our own design recommendation is therefore to collect only necessary data, prefer local processing, and enable a consultation without a photo. [3]

If multiple images of the same person appear in both training and testing, a model can use recognizable features. The test then appears more independent than it is. Similar problems arise if the subsequent target information is implicitly contained in the inputs already.

For fair evaluation, individuals and, if possible, recording conditions are cleanly separated. Additionally, an external test with new equipment or another location should take place. The accuracy must match the actual task: centimeter errors of a body measurement, classification accuracy of a position, and subsequent comfort are different results. A single high percentage cannot replace these differences.

A photo algorithm recommends H3 with a seemingly 98 percent certainty. Without a definition of the target value, this number is hardly interpretable. Does it mean agreement with previous salesperson judgments, with a product category, or with long-term comfort?

STOLL should translate the recommendation into verifiable features. Which body region likely requires more compliance, which models are available for selection, and what uncertainty remains? A subsequent lying test can confirm or correct the hypothesis. An exact number is only helpful if its meaning and calibration are proven.

Feedback after purchase can improve a system, provided it is collected voluntarily, structurally, and in compliance with data protection. Complaints and cancellations should be considered just as much as satisfied customers. Otherwise, the model learns from a biased selection.

Changes to the product range, covers, or manufacturing processes can change the assignment. A model therefore requires versioning and re-testing. For STOLL, a transparently maintained assistance system is more plausible than a supposedly universal body formula trained once.