logo SBA

ETD

Digital archive of theses discussed at the University of Pisa

 

Thesis etd-05302019-162625


Thesis type
Tesi di laurea magistrale
Author
MELONI, ENRICO
email address
enrico.meloni@outlook.it
URN
etd-05302019-162625
Thesis title
Using Virtual Worlds to Train an Object Detector for Personal Protection Equipment
Department
INGEGNERIA DELL'INFORMAZIONE
Course of study
COMPUTER ENGINEERING
Supervisors
relatore Prof. Amato, Giuseppe
correlatore Prof. Falchi, Fabrizio
correlatore Gennaro, Claudio
correlatore Dott. Di Benedetto, Marco
Keywords
  • deep learning
  • domain adaption
  • object detection
  • personal protection equipment
  • safety equipment detection
  • transfer learning
  • virtual dataset
Graduation session start date
21/06/2019
Availability
Full
Summary
Neural Networks are an effective technique in the field of Artificial Intelligence and in the field of Computer Vision. They learn from examples, without programming into them any previous knowledge. Deep Neural Networks saw many successful applications, thanks to the huge amount of data that is available with the growth of the internet.
When annotations are not available, images are manually annotated introducing very high costs. In some contexts, gathering valuable images could be impractical for reasons related to privacy, copyright and security. To overcome these limitations the research community has taken interest in creating virtual worlds for the generation of automatically annotated training samples. In previous works, using a graphics engine for augmenting a training is shown to be a valid solution.
In this work, we applied the virtual environment to approach to a not yet considered task: the detection of personal protection equipment. We developed V-DAENY, a plugin for GTA-V. With it, we generated over 140,000 automatically annotated images in several locations with different weather conditions. We manually annotated two real datasets for testing. We trained a network with this approach and evaluated its performances. We showed promising results: after training with only virtual data, the network achieves 51.8 mAP on real data and 87.2 mAP on virtual data. After applying Domain Adaptation, the network achieves 76.2 mAP on real data and 73.3 mAP on virtual data.
File