Ask HN: What vision model would you build if you had infinite data?

#1

Post by artem_sbx » Fri, Dec 18, 2020, 7:13 PM UTC

We've been building deep learning vision models in simulation, without any real labeled data... and it's working!

Have you recently worked on a vision model for which data was very difficult to source?

* Could be a rare event that happens once a week?

* Could be too difficult to get precise labels? (6D-pose, cuboids)

* Could be a budget & speed of iteration issue..

Really looking to broaden our understanding of applications beyond robotics & manufacturing.

Here is a recent example from our test bench (30 sec): https://youtu.be/D33acz5mI10

* Mask R-CNN trained on 10,000 samples from SBX Quicksynth with 3D scans from YCB Benchmarks, running on RealSense SR305 RGB channel.

* 3D scans from: https://ycbbenchmarks.com/

* Data from: https://sbxrobotics.com, built on UE4

Re: Ask HN: What vision model would you build if you had infinite data?

#2

Post by plaidfuji » Fri, Dec 18, 2020, 7:31 PM UTC

Possibly not relevant to your current simulation environment, but if your tech is generally focused on extrapolating from simulation data, I think there’s a market for this kind of thing in document parsing. Extracting tables, charts, text, etc from images, PDFs... paired with a document simulator could be powerful.

Re: Ask HN: What vision model would you build if you had infinite data?

#3

Post by brianjunyinchan » Sun, Dec 20, 2020, 12:07 AM UTC

Not involved in this area myself, but maybe some application of wildlife monitoring?

Re: Ask HN: What vision model would you build if you had infinite data?

#4

Post by alphagrep12345 » Wed, Dec 23, 2020, 3:45 AM UTC

plaidfuji wrote: post #2

Possibly not relevant to your current simulation environment, but if your tech is generally focused on extrapolating from simulation data, I think there’s a market for this kind of thing in document parsing. Extracting tables, charts, text, etc from images, PDFs... paired with a document simulator could be powerful.

Why would we need document simulators for it? Isn't all we need good OCR systems?