This repository aims to train a convolutional neural network with generated images of humans and monsters. The dataset is annotated with human body dimension. Monsters have dimensions which are "unnatural".
The images have a grey background. We first introduce artificial backgrounds to the dataset. Three datasets are then created:
- the original dataset with a grey background
- images with a texture background
- images with a random single color RGB background
The goal is to calculate their human body dimensions. Monsters and humans should be differentiated.
Human Body Dimension (HBD) Estimation using images becomes increasingly important. Datasets consisting of images of synthetically created human body meshes have been successfully used for training convolutional neural networks (CNN) to estimate HBDs. However, those images all have a uniform grey background which is rather improbable in practical application. But adding an artificial background to the image might make it harder to clearly identify the human shape. In this work, we introduce uniformly colored and textured background into an existing dataset of images of human body meshes. We develop and train an existing network architecture and then let it predict the HBDs. Our modified network achieved better results (∼50% less relative percentage error) than those reported in the original work we focused on. And it even accomplish remarkable results with different backgrounds.
