They propose Meta-Sim that generates a large dataset of labeled data for applications such as self-driving by generating the corresponding image from a document (label) of graph structure. The optimization is achieved not only by the loss of measuring whether the generated images correspond to the document, but also by the loss of optimizing the task (e.g., semantic segmentation).
TL;DR
They propose Meta-Sim that generates a large dataset of labeled data for applications such as self-driving by generating the corresponding image from a document (label) of graph structure. The optimization is achieved not only by the loss of measuring whether the generated images correspond to the document, but also by the loss of optimizing the task (e.g., semantic segmentation).
Why it matters:
Paper URL
https://arxiv.org/abs/1904.11621
Submission Dates(yyyy/mm/dd)
Authors and institutions
Amlan Kar, Aayush Prakash, Ming-Yu Liu, Eric Cameracci, Justin Yuan, Matt Rusiniak, David Acuna, Antonio Torralba, Sanja Fidler
Methods
Results
Comments