Synthetic environment for machine learning experiments
Mithun Lal · Queensland University of Technology · 2022
This thesis addresses the problem of data scarcity in human deep-learning applications. Automated estimation of human shape and pose from an image is challenging. It is even more difficult to map the identified human pixels onto a 3D model. Existing deep-learning models learn to map manually labelled human pixels in 2D images onto human surface, which is prone to human error, and the sparsity of annotated data leads to sub-optimal results. We solve this problem by generating realistic artificial human video data to train 2D-3D human mapping models and show promising results when compared to models trained on real data.