NCSI: Photorealistic Synthetic Data Generation for Deep Learning Training

Status	Completed
Mentor Name	Hao Ji
Mentor's XSEDE Affiliation	Education Allocation
Mentor Has Been in XSEDE Community	4-5 years
Project Title	Photorealistic Synthetic Data Generation for Deep Learning Training
Summary	Using synthetic data in deep learning training helps to reduce manual annotation, but often fails to produce satisfactory predictions on real data. One of the primary factors is the lack of visual realism the presence of objects in training images. Recent progress in image synthesis and adversarial generative networks have catalyzed the rapid generation of photorealistic images with limited user assistance. We plan to study an efficient approach to generating photorealistic, synthetic data and analyze the reality gap of synthetic data in deep learning.
Job Description	Students will use graphics rendering tools for creating photorealistic synthetic images and use them to train and benchmark deep learning models.
Computational Resources	XSEDE Bridges GPUs
Contribution to Community
Position Type	Intern
Training Plan	I will teach the students with the necessary knowledge and skills about realistic image generation, deep learning, and how to use XSEDE for training. During this project, I will work closely with the students and plan to have weekly meetings with them to guide their research progress.
Student Prerequisites/Conditions/Qualifications	Must have an undergraduate at California State Polytechnic University, Pomona
Duration	Semester
Start Date	01/18/2020
End Date	05/08/2020