A 3D Body from Eight Questions – No Photo, No GPU
3 days ago
- #anthropometry
- #3D body reconstruction
- #machine learning
- A 3D body can be generated from an 8-question questionnaire using a small MLP with physics-aware loss, achieving height accuracy of 0.3 cm and mass accuracy of 0.3 kg on CPU in milliseconds.
- The method improves on photo-based pipelines for circumferences without needing photos, addressing privacy, speed, and cost concerns, and avoids user photo-taking hassle.
- Initial inspiration came from Bartol et al.'s regression using height and weight, but limitations were found due to inability to capture body shape variations like build, belly, shape, and cupsize.
- Additional features such as build, belly, cup size, and gender were analyzed and found to significantly reduce waist variation, with build having the most impact.
- A dataset of tens of thousands of synthetic bodies was generated using the Anny model, focusing on 58 relevant blendshapes, with separate models trained for each gender.
- The MLP uses a loss function incorporating forward pass through Anny to ensure accurate height and mass predictions, addressing mass calculation issues that ridge regression failed on.
- Results show mean errors for height (0.3 cm), mass (0.3-0.5 kg), and BWH circumferences (3-4 cm), outperforming both Bartol's regression and photo-based methods on validation sets.
- Lessons include fixing inconsistencies in Anny's mass calculation by adjusting body density based on gender and body fat, and adding ancestry features to reduce noise and improve accuracy.
- Future improvements may involve interactive body adjustment features and incorporating more attributes like limb lengths, moving beyond static questionnaires for better user input.