Vision-based body measurements : Optimization of prediction accuracy
Väinölä, Jori (2021)
Väinölä, Jori
2021
Tieto- ja sähkötekniikan kandidaattiohjelma - Bachelor's Programme in Computing and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2021-03-23
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202102242240
https://urn.fi/URN:NBN:fi:tuni-202102242240
Tiivistelmä
Recovery of 3D body information from 2D images is important yet challenging task with many applications in industrial design, clothing and online shopping and medical diagnosis. The development in this field is however increasingly slowed down by lack of publicly available data-set of realistic 3D human models. In order to avoid this issue, silhouette models are used, which are generated with a parametric body model learned from real 3D body scans.
The purpose of this study to evaluate which parameter influences the result most in predicting body measurements from 2D images. The tested parameters were the resolution of the images, the amount of the generated silhouette models, type of neural network, direction of image in relation to the body, and number of images per silhouette model. By studying these factors, the goal is to find the optimal combination of the parameters in order to best optimize the accuracy of
the prediction model.
Results indicate that accuracy is highly correlated with training data size and functions best in combination with a simple convolutional neural network. Quality of the images, image angle, and use of multiple images also provide benefits, but the effect is smaller compared to network and data size.
The purpose of this study to evaluate which parameter influences the result most in predicting body measurements from 2D images. The tested parameters were the resolution of the images, the amount of the generated silhouette models, type of neural network, direction of image in relation to the body, and number of images per silhouette model. By studying these factors, the goal is to find the optimal combination of the parameters in order to best optimize the accuracy of
the prediction model.
Results indicate that accuracy is highly correlated with training data size and functions best in combination with a simple convolutional neural network. Quality of the images, image angle, and use of multiple images also provide benefits, but the effect is smaller compared to network and data size.
Kokoelmat
- Kandidaatintutkielmat [7049]