The foot is a vital human body part comprising a complex system of muscles and bones sustaining the human weight, and providing balance and mobility when daily activities are being performed. Extracting accurate foot measurements is of paramount importance in many applications including medical sciences, sports and fashion industry. Traditionally, footwear brands employ contact-based foot measuring methods involving a trained operator to design and produce well-fitted footwear products. However, this process is very time consuming and is prone to human errors. With the advancement of 3D scanning technologies, the foot can be scanned accurately with an affordable 3D scanning device. In this research, we propose, to the best of our knowledge, the first deep neural network (FNet) for automatic foot measurement extraction from a 3D foot point cloud. The proposed FNet is an encoderdecoder neural network which operates on the foot point cloud and outputs the foot reconstruction as well as the corresponding measurements points utilized for measurement extraction. Our study shows that teaching the network to accurately generate the measurement points, performed with the help of the well-designed loss functions, is necessary for automatic and accurate foot measurement extraction. In order to train the proposed neural network, a large dataset of complete foot scans with their corresponding measurement points and measurement values are synthesized. The performance of the proposed method has been evaluated on both synthetic test data as well as the real scans captured by the Occipital Structure Sensor Pro. The results show that our method outperforms the state-of-the-art methods in terms of accuracy and speed.

3D human body models are widely used in human-centric industrial applications, including healthcare, fashion design, body biometrics extraction, and computer animation. Prior to processing and analyzing body models, it is significantly important to rotate them to the same orientation. For instance, body measurement systems and virtual try-on systems usually assume the orientation of the body is known. These systems will output incorrect results or report errors without the correct orientation information. Unfortunately, the orientations of scanned bodies are different in practice since they are in different coordinate systems due to the setup variations of scanners. To automatically normalize the orientations of bodies is a challenging task due to the presence of pose variations, noises, and holes during 3D body scanning. In this study, we propose a novel deep learning-based method dubbed OrienNormNet for normalizing the orientation of 3D body models. As shown in Figure 1 and Figure 2, OrienNormNet directly consumes raw point clouds or mesh vertices, and it is applied in an iterative manner. First, the centroids of point clouds are translated to the origin to obtain zero-centered point clouds. Next, OrienNormNet consumes zero-centered raw point clouds and outputs coarse axis rotation angles. Finally, OrienNormNet takes the coarsely rotated point clouds from the previous processing as an updated input and outputs the refined axis rotation angles. By applying the obtained coarse and fine axis rotation angles, thousands of bodies can be adjusted to the same orientation in a few seconds. Experimental results based on synthetic datasets as well as real-world datasets validated the effectiveness of our idea.