Fitting Fashion Using Machine Learning

SAXION.NL/AMI Jeroen Linssen + Remco Booij + Adrian Brezoi + Bodhi Mulders + Matthijs van Veen Online shopping for clothes can be hard, because you cannot directly try on apparel. Because of this, webshops offer the service to take in returns when clothes turn out not to fit. For some shops, this is half of their total shipped clothing items, resulting in tremendous costs. This calls for solutions to help prevent such problems. Our goal: an online tool which can estimate a person’s body measurements based on 2D photographs of that person. In the tool, you can upload your pictures, after which your sizes are determined and a 3D model of your body is generated. Our approach relies on a model generated through machine learning. We use a convolutional neural network (based on AlexNet) that is trained on a dataset of 2D images based on 3D scans of people with their corresponding measurements. Improving speed and accuracy of FF-net and 3D model generation Rendering clothes on the virtual character Animating characters VR/AR support Online ClothesShopping Machine Learning for Size Estimation Training Neural Networks Open Challenges FF-net Dataset: 4012 silhouettes + measurements Training: Decrease in percentual error rate Deep learning network: FF-net Predicted size