Towards Fast and Robust Body Measurements Extraction
Mihai Petre, Cosmin Ciocîrlan, Eduard Cojocea, Traian Eugen Rebedea · 2022
parameters, such as weight, body circumferences, body fat percentage, metrics used by cardiologists and nutritionists.Moreover, MRI and RX scanners can be optimized, using the measurements in the calibration stage.Accurate human body measurements extraction is a complex problem, tackled in Computer Vision research since late 20 th century.The goal in this problem is to extract measurements with a maximum of 5 millimeters (mms) errorthe widely-accepted tolerance in tailoring.In the following sections we present an Ensemble Solution, which makes use of traditional Computer Vision techniques, as well as Deep Learning and Statistical Models. RELATED WORKOur solution uses techniques such as pose-estimation, semantic segmentation, and depth-estimation.In the following section, the state-of-the-art in these areas will be described.EfficientNet [1] is a convolutional neural network architecture and scaling method.It can scale depth, width, and resolution uniformly by using a compound coefficient.The coefficients are different from the conventional scales, as they are fixed.The logic behind the compound part is that, if the image is bigger, then the network will need more layers and channels to capture patterns.The compound scaling method can be generalized to an existing CNN architecture.Choosing a good baseline network is a priority, as the method only enhances the predictive capacity of the base network.The EfficientNet-B0 is based on the inverted bottleneck residual blocks of MobileNetV2 [2].PyTorch3D [3] is a framework from Facebook Research that handles working with meshes and it is designed to integrate with deep learning to predict and manipulate 3D data.This framework will be used to develop a version of the measurement prediction model, where it will receive the weight, height and the 2 images to generate the mesh of the person, from which more than 200 measurements can be extracted.In the next few paragraphs, some relevant history of the evolution of anthropometric features extraction will be presented.