3DBodyTech2025 - Improved Body Measurement Accuracy when Scanning Loose-Clothed Subjects скачать в хорошем качестве

3DBodyTech2025 - Improved Body Measurement Accuracy when Scanning Loose-Clothed Subjects

Scan-over-Clothes (SOC): Improved Body Measurement Accuracy when Scanning Loose-Clothed Subjects - 25.04 Extract of Presentation of Matthew S. GILMER, Size Stream at 3DBODY.TECH 2025 Full paper at https://proc.3dbody.tech/abstracts/20... Full video available in the proceedings Abstract: At Size Stream we continuously strive to enhance the accuracy of our 3D body scanning technology while minimizing user friction. We have previously employed a model that detects loose clothing during scanning. In the cases of positive detection, we have prompted the user with a suggestion to change into tighter-fitting attire. We now introduce a method that effectively compensates for loose clothes for our avatar generation and measurement predictions. For scans of subjects wearing clothing that is marginally loose and only partially covering the body, this compensation results in an accuracy degradation of less than 10% compared to scans of those same subjects wearing ideal scanning attire. This improvement was achieved through two primary developments. First, we developed a Body-Plus-Clothes (BPC) segmenter which, in addition to separating the subject from the background, distinguishes the clothing from the bare body. Aside from the added capability to segment clothing, this model also provides a substantial accuracy improvement in separating the subject from the background when compared to our previous segmenter (S2023). The BPC segmenter was trained using real images and supervised to manually-corrected output from an open-source segmentation model. Second, leveraging the BPC segmenter, we developed a Scan-Over-Clothes (SOC) Body Measurement Model (BMM) specifically designed to adjust for the presence of loose clothing during body reconstruction and measurement estimation. The SOC model was trained using a combination of BPC segmentation of real images and internally-generated synthetic data. This novel approach results in a solution that substantially enhances measurement accuracy for scans involving loose clothing, raising the possibility for reliable Ready-To-Wear (RTW) sizing with relaxed scan wear requirements.