4 ms·
I don't disagree that hardware itself has more dependencies. I don't think that makes it harder or less probable of delivering though than bleeding edge compute
by AndrewKemendo 10y ago
I don't disagree that hardware itself has more dependencies. I don't think that makes it harder or less probable of delivering though than bleeding edge computer vision/machine learning systems.
I too have worked hardware and on those projects we could get fairly low level line workers on assembly up to speed quickly, source our parts in a repeatable fashion and build systems to scale without too much brain power. Not easy, but more logistics management than creativity.
With these non-deterministic CV systems (for example point cloud generation) the path to working is much less clear.
I'm not saying one is harder than the other, but the class of unsolved problems in CV/ML doesn't have off the shelf solutions, so they pose different but similarly hard problems as hardware.
- x0x0 10y agoI still think ml systems, which I work on, are easier than hardware, which I used to work on. It's things like the cost to dump inventory if you make a mistake (eg a founder cheaped out on a ballast circuit on a dna reader, which caused the light to flicker at power on, which fucked up the chemistry, which cost the company well over a million dollars), etc. Which is not to imply novel ml systems are easy... Pair3d looks really cool -- I've been making a list of neat industrial uses of computer vision tech and virtual showrooms isn't something I'd thought of. A friend actually did something similar the hard way -- he and his fiance were apartment shopping in nyc, but the apartments were empty and they were having a hard time visualizing what they would be like (less spacious, for starters) with furniture. So they cut butcher paper into pieces the size and shape of their major furniture and laid it out in the apartment to see what it would feel like with their couch, bed, etc in the apartment rather than empty.
- snovv_crash 10y agoA lot of the CV problems are essentially 'solved', though. Feature extraction (AKAZE, BinBoost), matching (GPU brute force hamming), RANSAC (with PROSAC and relatives), bundle adjustment (Schur Decomposition with LM), point clouds (SGM, PatchMatch) and mesh (Poisson, FSSR). Each stage has tuning, and real-time requires sacrifices on the hardware we have available, but we know with better computers we can have it. (HSA has me drooling, hurry up with Zen, AMD!) IMO the harder stuff is in semantic segmentation of point clouds and dynamic scenes, but I have high hopes for the next few years.
- AndrewKemendo 10y agosemantic segmentation of point clouds lol, hey it's hard enough to do with static images. Feature matching pointclouds is probably turing complete :P. That's the kind of shit we're working on though. We're trying to turn the real world into a platform.
- snovv_crash 10y agoI feel like it should be easier than images, since we have all the 3D information. There are all of the features based on 3D structure which image segmenters can't even begin to use. It's one of those things that we can do efficiently, so with enough priors of what scenes look like a sufficiently informed ML system should be able to get decent accuracy.