Iqra Nosheen

Project Title

There have been notable successes in Deep Learning, but the requirement to have large, annotated datasets creates bottlenecks. Datasets must be carefully compiled and annotated with ground-truth labels. One emerging solution is to use 3D modelling and game engines such as Blender or Unreal to create realistic virtual environments. Virtual cameras placed in such environments can generate images or movies, and since the locations of all objects in the environment are known, we can computationally generate fully accurate annotations. Drawing on the separate and complementary fields of experience of the two supervisors, the PhD student will gain a synergy of expertise in Graphics Perception and Deep Learning. This PhD research will investigate questions including: (1) strategies to combine real-world and virtual images; (2) the importance of realism in virtual images; (3) how virtual images covering edge cases and rare events can increase the reliability, robustness and trustworthiness of deep learning.

Supervision Team

There have been notable successes in Deep Learning, but the requirement to have large, annotated datasets creates bottlenecks. Datasets must be carefully compiled and annotated with ground-truth labels. One emerging solution is to use 3D modelling and game engines such as Blender or Unreal to create realistic virtual environments. Virtual cameras placed in such environments can generate images or movies, and since the locations of all objects in the environment are known, we can computationally generate fully accurate annotations. Drawing on the separate and complementary fields of experience of the two supervisors, the PhD student will gain a synergy of expertise in Graphics Perception and Deep Learning. This PhD research will investigate questions including: (1) strategies to combine real-world and virtual images; (2) the importance of realism in virtual images; (3) how virtual images covering edge cases and rare events can increase the reliability, robustness and trustworthiness of deep learning.

Description

There have been notable successes in Deep Learning, but the requirement to have large, annotated datasets creates bottlenecks. Datasets must be carefully compiled and annotated with ground-truth labels. One emerging solution is to use 3D modelling and game engines such as Blender or Unreal to create realistic virtual environments. Virtual cameras placed in such environments can generate images or movies, and since the locations of all objects in the environment are known, we can computationally generate fully accurate annotations. Drawing on the separate and complementary fields of experience of the two supervisors, the PhD student will gain a synergy of expertise in Graphics Perception and Deep Learning. This PhD research will investigate questions including: (1) strategies to combine real-world and virtual images; (2) the importance of realism in virtual images; (3) how virtual images covering edge cases and rare events can increase the reliability, robustness and trustworthiness of deep learning.