MSc Thesis Defense by: Kavan Mehulkumar Dave

Computer simulation is a powerful technique to create synthetic images for the testing of computer vision algorithms. However, the discrepancies caused by inconsistencies between the simulated environment and the physical world have made the results of algorithms testing with synthetic datasets hardly reliable in real-world applications, which is a phenomenon called the “reality gap”. Among various factors that impact on photo-realism of computer-synthesized outdoor environments, researchers have analyzed the most influential and identified them as geometry, appearance, lighting, physics, environment, camera, and rendering parameters.

 

This thesis research develops a novel system for the generation of synthetic datasets with a reduced reality gap. In the system, a scene environment is first constructed with the geometry of objects according to real-world data. Information retrieved for the time of a day from the Internet of Things (IoT) is then used by a simulator engine to render images with the actual appearance of objects, physics of shadows and reflections, and rendering parameters, such as camera, lighting, and environmental conditions. Similarity scores between the synthesized and real-world images are finally calculated to evaluate the effectiveness of the proposed system in reducing the reality gap.

 

Keywords: Computer Vision, Simulation, Synthetic Data, IoT

 

Internal Reader: Dr. Imran Ahmad           

External Reader: Dr. Mohammed Khalid

Advisor: Dr. Xiaobu Yuan

Chair: Dr. Dan Wu