Waymo Releases a Self-Driving Open Data Set For Free Use By Research Community (techcrunch.com) 10
An anonymous reader quotes a report from TechCrunch: Waymo is opening up its significant stores of autonomous driving data with a new Open Data Set it's making available for the purposes of research. The data set isn't for commercial use, but its definition of "research" is fairly broad, and includes researchers at other companies as well as academics. The data set is "one of the largest, riches and most diverse self-driving data sets ever released for research," according to Waymo principal scientist and head of Research, Drago Anguelov, who was at both Zoox and Google prior to joining Waymo last year. Anguelov said in a briefing that the reason he initiated the push to make this data available is that Waymo and several other companies working in the field are "currently hampered by the lack of suitable data sets."
The Waymo Open Data set tries to fill in some of these gaps for their research peers by providing data collected from 1,000 driving segments done by its autonomous vehicles on roads, with each segment representing 20 seconds of continuous driving. It includes driving done in Phoenix, Ariz.; Kirkland, Wash.; Mountain View, Calif.; and San Francisco, Calif., and offering a range of different driving conditions, including at night, during rain, at dusk and more. The segments include data collected from five of Waymo's own proprietary lidars, as well as five standard cameras that face front and to the sides, providing a 360-degree view captured in high resolution, as well as synchronization Waymo uses to fuse lidar and imaging data. Objects, including vehicles, pedestrians, cyclists and signage is all labeled. "We decided to contribute our part to make, ultimately, researchers in academia ask the right questions -- and for that, they need the right data," Anguelov said. "And I think this will help everyone in the field; it is not an admission in any way that we have problems solving these issues. But there is always room for improvement in terms of efficiency, scaleability, amount of labels to need. It's a developing field. Mostly we're trying to get others into thinking about our problems and working with us, as opposed to doing work that's potentially not so impactful, given the current state of things."
The Waymo Open Data set tries to fill in some of these gaps for their research peers by providing data collected from 1,000 driving segments done by its autonomous vehicles on roads, with each segment representing 20 seconds of continuous driving. It includes driving done in Phoenix, Ariz.; Kirkland, Wash.; Mountain View, Calif.; and San Francisco, Calif., and offering a range of different driving conditions, including at night, during rain, at dusk and more. The segments include data collected from five of Waymo's own proprietary lidars, as well as five standard cameras that face front and to the sides, providing a 360-degree view captured in high resolution, as well as synchronization Waymo uses to fuse lidar and imaging data. Objects, including vehicles, pedestrians, cyclists and signage is all labeled. "We decided to contribute our part to make, ultimately, researchers in academia ask the right questions -- and for that, they need the right data," Anguelov said. "And I think this will help everyone in the field; it is not an admission in any way that we have problems solving these issues. But there is always room for improvement in terms of efficiency, scaleability, amount of labels to need. It's a developing field. Mostly we're trying to get others into thinking about our problems and working with us, as opposed to doing work that's potentially not so impactful, given the current state of things."
Useless - does not allow redistribution of models (Score:3)
Does not allow redistribution of models. How the fuck am I supposed to publish research, if I can't offer pretrained checkpoints so that others verify the results?
Re: (Score:2)
Does not allow redistribution of models.
Neither TFS nor TFA says that.
Why do you think Waymo prohibits redistribution of models?
Read the terms (Score:3)
It's in their terms that you have to agree to before you download.
Re: (Score:1)
I dunno. Maybe you publish the methods you used to train your models, mention which data sets you trained it on. So other researchers can, maybe, I dunno. Reproduce your work?
Sort of like you're meant to be able to do in a scientific paper.
-David
Re: (Score:2)
Before you're able to reproduce (and recent DL papers can require MONTHS to reproduce) it'd be pretty cool to at least measure the metrics. That is, reproduce the actual outcome of the research. Authors can (and do) put all sorts of non-reproducible shit in their papers and omit crucial details. It'd be nice to see if their research holds water in the first place before spending the effort to reproduce.
5.5 Hours? (Score:2)
Doesn’t seem like a whole lot of data, diverse circumstances though it may be. Curious how useful it really is.
Indicator of capitulation? (Score:2)
Is this a sign that autonomous driving is hard?