Democratizing object detection training: utilizing large scale simulations to train neural network models


Deep learning models specifically in the object detection task require tremendous amount of labelled images to achieve the superhuman performances. How can we get these labelled images? While there exist large open-source datasets, they are not exhaustive and they cannot be customized to the requirements. Large scale high-fidelity simulator such as CARLA which utilize gaming engine to generate realistic simulations can be used to generate infinitely complex datasets with perfect ground truth without any user intervention. The challenge then lies in making the model robust to real data. We propose to develop novel object detection models which can be used for object detections in naturalistic datasets with minimal human effort. Our goal would be to utilize these models as a workhorse in future external projects and collaborations.

People

Paul
Green

IOE, UMTRI
Engineering

Carol
Flannagan

UMTRI
Engineering

David
LeBlanc

UMTRI
Engineering

Jim
Sayer

UMTRI, CEE
Engineering

Arpan
Kusari

UMTRI
Engineering

Wenbo
Sun

IOE, UMTRI
Engineering


Funding

Funding: $30K (2022)
Goal: Develop novel deep learning based object detection models trained on high fidelity simulators and employed on real datasets.
Token Investors: Paul Green, Carol Flannagan


Project ID: 1027