Hacker News Reader

Comment by amelius | Hacker News Reader

The author could also have used 3d simulations to obtain more data, and determine feasibility and proper hyperparameters for the network before training with actual data.

View on HN

3D simulations can’t create photorealistic images. (Not currently, anyways)

You don't need photorealistic images for these situations. The images are always downsampled to a smaller resolution. So getting "good enough" but not photorealistic is fine.

If you have software to simulate real-world physics for hundreds of independent agents that can be run on a personal computer with close to real world graphic fidelity (including realistic noise), you’re running a multibilion dollar company, not writing thought pieces on Medium.

There are a few companies you can get simulation software from specifically designed for self-driving cars. See https://blogs.nvidia.com/blog/2017/10/26/self-driving-simula... for some discussion.

It's true that this isn't photo-realistic (see https://www.youtube.com/watch?v=D7OccC0W_CU for a video), but at the same time it is useful enough for training.

There were a couple of projects using GTA 5 for training data successfully (one from OpenAI) but they got taken down due to licensing issues

One option is to train a GAN to do domain adaptation. Google (disclaimer: I work there) recently did that very successfully for training robots to grasp: https://research.googleblog.com/2017/10/closing-simulation-t...

Reread my comment. I didn't suggest to only use the simulations.

Check out "domain adaptation" https://en.wikipedia.org/wiki/Domain_adaptation

Reply on news.ycombinator.com