Year: 2017
Type: dissertation
Abstract: The purpose of this master thesis was to evaluate how a deep convolutional neural network can be used to estimate a human body pose in a 2D image.The evaluation was done by implementing a state of the art network and test its performance on the MPII human pose dataset.Three different ways to preprocess the training data was tested and the network's accuracy went from 60.8% for the first method to 71.2% for the third method, a roughly 17% increase in performance.These results indicate that the choice of preprocessing method have a large effect on the overall performance of the network.An analysis of the MPII dataset revealed that the distribution of poses is skewed towards poses facing the camera in an upright position.A score was calculated for each pose that measures how much a pose deviates from the norm.Using this pose score it is shown shown that the network performs much worse on poses that are rare in the MPII dataset.This indicates that network has learned how the keypoints are distributed and has overfitted to the MPII dataset.To improve the generalization of the network a more thorough data augmentation should be implemented.
Author: Sven Abelsson Runing
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2780838738