AIS-Bonn / stillleben

Physics-based Scene Synthesis for Deep Learning
https://ais-bonn.github.io/stillleben/
MIT License
60 stars 2 forks source link

Image Generation Speed #1

Closed themasterlink closed 4 years ago

themasterlink commented 4 years ago

How many different images can stillleben produce per second?

xqms commented 4 years ago

That highly depends on your requirements. The costly operation is physics simulation (i.e. scene generation), which can take up to 100ms depending on the scene complexity. However, you can render many different views from one object arrangement using random viewing poses. Additionally, if you properly pipeline the scene generation using multiple processes, you can easily achieve >100 fps on a single GPU. I think it could be pushed even further, but I never needed higher rates so far - CNN training is much slower ;)

The YCB example we provide is simplified as much as possible and does not use pipelining and multi-process computation. I'll add a more complicated example optimized for performance soon.

themasterlink commented 4 years ago

Thank you very much, that answers my question!