Fast and Accurate 3D Foot Reconstruction from a Single Image

Obtaining accurate 3D reconstructions of the human foot from 2D images holds significant importance in various fields, including anthropometry, footwear design, and medical diagnostics.In this study, we propose a novel neural network-based approach for reconstructing a 3D mesh of the foot from a single image.Our method integrates multiple sources of information, including binary segmentation masks and 2D keypoint estimation.By leveraging Principal Component Analysis (PCA) to represent foot morphology in a low-dimensional space, we infer the parameters needed for 3D mesh reconstruction, including rotation and translation parameters for alignment with the input image.Our approach builds upon recent advancements in deep learning for 3D reconstruction from images and demonstrates promising results in accurately capturing foot morphology.The model has been trained with two datasets: one consisting of 1M synthetic samples and another with 500K augmented real samples.Validation on a test subset of over 674 samples resulted in a PA-MPJPE (Procrustes-Aligned Mean Per Joint Position Error) of 0.9 mm.Furthermore, the real-time capability of our method makes it suitable for applications in augmented reality, such as virtual try-on and improvements in user experience and precision of phone-based foot scanning solutions.