Once Dataset: One Million Scenes for Autonomous Driving
once-for-auto-driving.github.io
once-for-auto-driving.github.io
Damn it would be cool if I could rank this data set too...
How does that work? What kind of data is being sold and who's buying?
What kind of data?
Check on https://rugpullindex.com for a summary ("Data Set" column)
Who is buying?
There's speculation. I don't think there's any real consumption yet.
Blockchain tech in a nutshell.
> Don't be snarky.
> Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize. Assume good faith.
Maybe that's reasonable formulation of their opinion, and you shouldn't assume it's snark.
Speaking as a jaded professional data scientist, I couldn't be more pleased about this development.
I tend to still post my project here on HN as I get tons of feedback and since HN's traffic is just enormous. Also I just don't have a lobste.rs account lol. On HN, I have my email ready for people to send me stuff. I get a fair amount of freelance offers. But I have to say that most of my HN usage is pretty discouraging.
I'm a self-employed freelancer and I try to work on what I find fun and allows me to code as much as I can. I do it because I enjoy it. However, often when I post to HN, I get a default treatment I honestly don't think I deserve. HN hates blockchain and HN hates so much other things for seemingly irrational reasons that I carry no guilt for and I cannot comprehend why there's so much hate.
Sure, I act in self-interest (e.g. by sharing my project or a blog article) so I expect a certain amount of scrutiny. But at least in my personal experience the negative feedback in comments overwhelms the positive and so using HN to talk about side projects has become super discouraging.
It's quite ironic, considering that it's probably ycombinator's goal to seed a community of people brave enough to build what they deem useful and valuable so that more investment opportunities arise.
Since months, I've now tried to seek communities that truly appreciate someone taking both a financial and reputational risk to build their side project publicly. I've not been too successful yet, unfortunately. If I could, I'd leave HN behind.
If anyone has communities I could fit in, please let me know. Also, invite me to lobste.rs pls thx.
It strikes me that this isn't nearly enough classes. Other important ones that I can think of offhand include deer (other animals too - recognizing the species helps with anticipating how they might move), playground ball, shopping cart, fallen limb, bicycle flag... I could possibly go on all morning.
Even looking at the list has me thinking that treating this as a simple classification subtask is possibly going about it in the wrong way. At least in my city, you need to be able to make finer distinctions, such as between a pedestrian stepping out into the street to see if a bus is coming or look for a taxicab, a panhandler standing on the centerline, or a pedestrian stepping out into the street to cross it. This is a legitimate safety concern for city driving - if a self-driving car hits the brakes every time it sees someone looking for a bus, it's only a matter of time until someone rear-ends it. And, while it should definitely give that panhandler some space, stopping in the middle of a busy intersection to let them clear the road when they have no intention of doing so is going to excessively obstruct traffic, and possibly also create a further safety hazard as cars behind it maneuver to try and get around.
I'm not sure figuring these things out is really a classification problem in the first place. It seems like it's more of a, for lack of a better term, "inferring intent" problem. Part of what I think might make that distinct is that classification assumes you're working with a categorical variable - a finite set of distinct classes. Intentions that are relevant to the driver tend to clump into a few distinct categories, but they're perhaps better modeled as fuzzy regions in a continuous vector space than as a categorical variable.
Which isn't to say that datasets like this are useless. Progress is progress, obviously. But it seems possible that, in the long view, the most important thing that all of this progress might have accomplished thus far is helping us to get out of the Dunning-Kruger phase of our understanding of the problem.
If this was published as one or more torrents it would be ideal: all of those who have downloaded the full set would be able to contribute bandwidth to help newcomers, and could spare the source from having to serve multiple TBs of content directly. Uploading to Archive.org and/or the Chinese equivalent would be even better, since they provide mirroring plus torrents plus direct downloads automatically and for free.
If you wanted to train, say, a distant traffic light recognizer, you could do that by training one to recognize close-ups of traffic lights. Then tag close traffic lights in video with it, track the tagged object back to when it first appeared, and that becomes a new tagged training case for distant traffic lights. Same for other objects.