kane1204/3d-pose-estimation-project

Part 1 of my Masters project

3D Monocular Pose Estimation from Synthetic Data

This is the first part of my masters project

Abstract

3D pose estimation has been a very popular part of computer vision for a long time now. The main challenge these days is to find data that can be used to train a 3D pose estimation algorithm since it requires 3 dimensional data. Our project addresses this problem by using a synthetic dataset based on the Sungaya stick insect. The dataset used is artificially generated to provide photo realistic samples. These samples are used to train our 3D pose estimation pipeline which consists of 2 deep neural networks. The first of these networks detects the 2D pose of a stick insect and the second network detects the 3D pose of the stick insect based of the 2D pose data from the previous network. The project also investigates how well these networks trained on synthetic data and carries out experiments on both models to determine how well the models are able to perform when faced with real data. Using the results provided by our investigations we determine that this pipeline is a good step into creating a 3D pose estimator for a stick insect which is able to detect 3D pose only trained on synthetic data.