Machine Learning
This paper introduces two innovations designed to reduce scan wear requirements for Size Stream’s mobile scanning platform: the Body-Plus-Clothes (BPC) segmenter, which distinguishes between a subject’s body and their clothing at the pixel level, and the Scan-Over-Clothes (SOC) Body Measurement Model, which uses BPC segmentation to compensate for loose clothing during body reconstruction and measurement prediction. Together, these advances are evaluated against existing algorithms to quantify the accuracy of trade-offs of scanning in loose versus tight-fitting clothing.
By reducing the accuracy degradation from loose clothing to under 10% for male subjects, and significantly less than what the prior model produced, Size Stream moves meaningfully closer to a scan experience that requires no special attire preparation, lowering a key barrier to adoption for both health monitoring and apparel applications. If extended successfully to ready-to-wear sizing, this technology could enable reliable garment sizing from a casual, fully clothed smartphone scan, fundamentally expanding the addressable market for mobile 3D body measurement.
Key takeaway
Size Stream’s Scan-over-Clothes (SOC) model is the first solution trained to recognize loose clothing and compensate for it, reducing accuracy degradation from everyday attire from 44–48% down to roughly 10%, while the underlying Body-Plus-Clothes segmenter delivers 12–14% better measurement variability even in ideal conditions. Trained on over 11,700 real-world images using a synthetic data approach, the model generalizes across garment types without requiring paired outfit scans for every subject. If SOC performs in ready-to-wear sizing trials as the accuracy results suggest, reliable garment sizing from a fully clothed, unprepared smartphone scan becomes a realistic near-term reality, a fundamental shift in how 3D body measurement can be deployed across retail, e-commerce, and consumer health.