For the past year I have been studying and building synthetic training
data with Houdini, a procedural 3D engine. It unexpectedly opens up
ways to partly solve problems that come up with ordinary synthetic
data generation.
The post describes all the mistakes I made, and the rules I ended up
with for closing the general problems synthetic data has.
I am a solo founder from Nizhny Novgorod. This is my first attempt at
a startup; my background is VFX and advertising.
I would be glad to talk to anyone who has done something similar —
what came out of it for you, what metrics you got, and how you solved
these problems.
This sounds very interesting. Was the gap between synthetic and real data significant? How did you deal with it? Have you tried transferring the technology to other objects?
Yes, there were a lot of different approaches, but the ones that really
worked were repeating the physics and watching the small things — some
extra reflection you would not think about sometimes works very well. And
the other way round, a beautiful picture can work worse than an ugly but
correct one.
The post describes all the mistakes I made, and the rules I ended up with for closing the general problems synthetic data has.
I am a solo founder from Nizhny Novgorod. This is my first attempt at a startup; my background is VFX and advertising.
I would be glad to talk to anyone who has done something similar — what came out of it for you, what metrics you got, and how you solved these problems.