Capturing Machine Learned 3D Foot Shapes from a Single Camera

点击翻译按钮获取中文摘要

In footwear, fit determines comfort and performance, and is highly dependent on foot shape, something that is not fully captured by shoe size. Affordable scanners based on RGBD cameras can be used to acquire more detailed sizing information, and allow for more personalized footwear matching. When scanning an object, many images from different views are usually required to reconstruct the overall shape, however prior information can be leveraged to more efficiently recreate models and fill in missing information. Deep learning methods have been shown to be able to reconstruct 3D shape from limited inputs in objects such as furniture and vehicles. This approach can be applied in 3D scanning, where a complete scan can be formed from a single input view. We apply a deep learning approach to foot scanning, and present a method to reconstruct a 3D point cloud scan from a single input depth map. Anthropomorphic body parts can be challenging compared to other objects studied in literature due to their irregular shapes, difficulty for parameterizing and limited symmetries. We will leverage MPII Human Shape models built from the CAESAR dataset to train a view synthesis based network. We will investigate using as few as one camera, either a RGB-D camera or just a RGB camera.