That's a pretty well-solved problem at this point, if you want to do it yourself. You'll want some kind of NeRF tool and a way to calculate the camera poses of the photos you took. COLMAP is the tool most people use for the latter.
I'd recommend trying Instant Neural Graphics Primitives (https://github.com/NVlabs/instant-ngp) from NVIDIA. It's a couple years old, so not state-of-the-art, but it runs on just about anything and is extremely fast.
Sweet, thank you for sharing. In my case, I need an api I can call cause i only have a mac air which is essentially worthless for development lol. Also I am bootstrapping a startup and one of the features is essentially turning rooms into 3d space. I know theres matterport 3d and some others but still looking for something simple where i could pay a couple cents per api call with x amount of images. does that make sense?
Statistics are not theft. Judges have written over and over again that training a neural network (which is just fitting a high-dimensional function to a dataset) is transformative and therefore fair use. Putting it another way, me summarizing a MLB baseball game by saying the Cubs lost 7-0 does not infringe on MLB's ownership of the copyright of the filmed game.
People claiming that backpropagation "steals" your material don't understand math or copyright.
You can hate generative tools all you want -- opinions are free -- but you're fundamentally wrong about the legality or morality at play.
I used to work in perception for autonomous aerial vehicles and horizontal wires were the hardest common object to avoid. Traditional stereo won't help you localize them -- wires are thin so even mere detection can be hard, and one portion of a wire looks much like another so feature matching fails resulting in bad or no depth estimates -- and LIDAR sacrifices resolution for weight and power consumption (which both have to be optimizied for drones). It's been years since I've worked in this field, and Amazon has many smart people thinking about it but I'm not surprised it's still a difficult problem.
> Traditional stereo won't help you localize them [...] and LIDAR sacrifices resolution for weight and power consumption
I wonder if a more mechanical solution wouldn't help:
Whiskers, like on a cat. A long enough set of thin lightweight whiskers could touch the wire before the propellers do, giving time for the drone to stop and change course. Essentially, giving the drone a sense of touch.
I hadn't thought about this in a long time. Looks like her lab is still going strong doing research at the intersection of biology and robotics on whisker-based sensing:
Thin lightweight whiskers are going to be challenging to manage on a propeller-driven vehicle. They'll get blown all over the place. Having them extend out past the propellers will likely get them tangled in the propellers.
But that's fine, isn't it? If they're intended to detect fixed objects, then noticing that one or more of them have ceased to be blown around in that way may be a good way to detect unanticipated contact with a fixed object: When the signal becomes less noisy, then maybe something is in the way.
And the whiskers don't have to be all floppy like a wet noodle. I myself am thinking that something rigid or semi-rigid might be good. Perhaps something akin to armature wire, or thin spring steel. Maybe even literal bamboo chopsticks.
They can also be constrained so that they don't get sent into the props.
My little brain thinks that the drone-end of the whiskers can be attached to potentiometers, with light return springs to bring them back towards center, like the mechanism used by an analog stick on a PS3 controller.
> And the whiskers don't have to be all floppy like a wet noodle. I myself am thinking that something rigid or semi-rigid might be good.
I don't think you're right about this. The concept of the whiskers is to notice when you've collided with something. Real whiskers aren't rigid because colliding with something when you're rigid means snapping. (Ever stub your toe?)
Think of the rigidity of the whiskers as being traded off against your maximum movement speed.
Suppose I've got an assembly with a chopstick attached to a gimbal with some minor centering springs and sensors (potentiometers) inside. The chopstick has many degrees of free angular movement provided by this gimbal and overall assembly.
I gently bounce ("slam"?) that chopstick off of a thing, and this results in the feedback loop that provides positioning control to provide immediate instruction to back off in the opposite direction of the apparent impact.
Does the chopstick take damage? Does the gimbal take damage? Does the greater assembly take damage?
Why, or why not?
(I feel like we're speaking two different languages here. Have you ever looked at how a PS3 analog stick works, or have you not? It's not new tech. It wasn't even new when it was new, and it's very nearly 20 years old now in PS3 form.)
Yes, you've successfully confirmed: We're quite clearly speaking different languages.
(Good luck with...whatever it is that you may be talking about. My diction is good. I don't have time or patience to explain it for outliers who aren't following along well and who also insist that it must somehow be wrong. I apologize for this; I am actually sorry.)
Rigid whiskers have other sets of problems. Below someone mentioned that rigid whiskers will break when they contact objects. If the whisker is as rigid as the drone itself, it plausibly breaks the same cables that the drone breaks. You also have the problem that in the event of drone failure, you now have a spike-covered drone falling out of the sky. What kind of damage does a bamboo chopstick or thin piece of steel do when it hits someone or something at ground level at drone-falling velocity with the mass of a drone behind it?
It's quite possible that these problems are solvable and can be engineered around, that there's a whisker-based solution, but I don't see it. It's certainly not an obviously workable solution.
A cage around the drone, there are kids toys like this, and also commercial products for inspection. Prevents contact with other objects, contact can be sensed and reacted to. https://www.flyability.com/elios-3
Doesn't protect against everything, like Spanish Moss which dangles from trees, but that is a lot bigger than a long thin wire.
Ah yeah I came up with the solution to that one. It's 'don't fly drones over our heads' approach. Also the 'upgrade the fragile infrastructure so a light breeze doesn't take out millions of people's power.'
It’s really hard for people too. The advice I got for landing in a field was to assume that every pole you saw had wires going to every other pole. Which is reasonable enough for that scenario, but not workable for continual low altitude flying in a built up area.
> horizontal wires were the hardest common object to avoid... Traditional stereo won't help you localize them
This makes a lot of sense. I wonder if it wouldn't be better for autonomous vision to use three cameras instead of two for better spatial reasoning.. maybe in a triangle pattern?
We experimented with a rig with more cameras on it (four, in a square) but the baseline of the cameras on the drones we were using could be measured in centimeters, so the vertical stereo pairs didn't provide much better results. Further, more cameras means more power, more weight, and much more expensive on-board processing (which also will require more power).
If you have a map of all utility poles you could probably just avoid every straight line between any of them within some reasonable distance of eachother.
it's an approximation of dangerous areas, catenary curves are more accurate than straight lines but you don't know the length of the cable so you don't know the droop height.
All cables? Everywhere in the entire country? Accurate to the centimeter level and updated on the hour?
Edit: This was flippant, but the real issues are: any map you get will be incomplete and obsolete almost immediately and cables move and sway in the breeze.
I used to work at Valve -- on the CS:GO team, no less -- although I left nearly a decade ago. I don't know what prompted this change but I have some suspicions. Even when I was there and the loot box system was new to CS:GO, there were concerns that a lot of trading was happening outside of the marketplace. The trading happened elsewhere because you can't have more than $300 in your Steam wallet (more than this would trigger some banking regulations that Valve wanted to avoid), so anything more valuable than that had to happen on 3rd party sites.
We didn't want this for three reasons: we'd lose out on the marketplace cut (10% of all sales I think?); we didn't want people grinding the game to earn money from rare drops; and finally because 3rd party trading ended up creating a lot of scams and therefore angry players.
At the time, we didn't see any way around it: we couldn't prevent people "gifting" items to each other, and despite omniscience and omnipotence in the game and Marketplace, we weren't confident that we could rejigger the drop rates and rarities to lower the maximum perceived value of the fanciest knife to be under the $300 limit.
I suspect that the CS:GO team finally decided to do something about it and chose this. If the team is anything like I left it, they probably modeled this extensively (we had data on nearly every game ever played in CS:GO and complete Marketplace data), and discussed the change with the TF2 and DOTA teams, who also have to deal with this, and decided that the short-term fury of a small fraction of the playerbase was worth it. I wonder if TF2 and DOTA are having similar problems and, if so, whether this change will be rolled out for those games, too.
Thank you for sharing your experience and perspective.
> we'd lose out on the marketplace cut (10% of all sales I think?); we didn't want people grinding the game to earn money from rare drops;
My naive understanding is that by having skins be worth tangible and significant value; this was the primary motivator for players to purchase keys to unbox cases, which was the dominant direct revenue generator for CS.
I would guess that the revenue generated from keys (and cases, from the market cut) eclipses the potential market cut revenue from limiting the value of items to the marketplace limit (now $2k I believe), as the consequence of that is significantly less demand in keys and skins as a whole.
Without the prospect of extremely expensive chase items, the $2.50 + ${case} slot machine pull loses its jackpot. With a knife being dropped once every 400~ unboxes, the EV of a knife would be $1000 + 400*${case}. Obviously the actual EV would be lower in practice, but the point I'm trying to understand is how the monetization model works if skins are any less expensive than they were.
> My naive understanding is that by having skins be worth tangible and significant value; this was the primary motivator for players to purchase keys to unbox cases, which was the dominant direct revenue generator for CS.
Yes. The Valve philosophy on the cosmetics marketplace (we called it "the economy") is that you distribute random rewards to players and they can trade and sell and discover the value of those goods for themselves. Obviously, this was done to make money for Valve but, in theory, it's also good for the players. It allows people who have things they don't want to sell them to people who want them. And all this buying and selling happens between Steam wallets (and there is no off-ramp) so at the end of the day, it's all just profit for Valve.
But above all we wanted people to play CS:GO because it was a fun game. We didn't want to turn it into some kind of grim pachinko parlor, with players grinding out matches just to get random loot box drops. So you have to balance the potentially real dollar random rewards so that they're a fun surprise but not economically attractive enough to become a job.
It’s just a weird side effect that’s surprisingly difficult to prevent - online games have had gold farmers for pretty much as long as there have been online games with gold.
I’ve run into idle bot accounts several times while playing and it’s infuriating. Mainly in the arms race mode.
Players can leave and join that mode at any time. So the bots will constantly be joining and leaving. if the bots manage to become 50% of the game they will vote kick all the remaining players. I’ve had several in progress matches interrupted because a few of the actual players bailed and the bots managed to take over the lobby.
Runescape tried this back in 2007 along with completely disabling PvP; it was a very unpopular change for the vast majority of players who were not buying items.
I stopped playing the game around that time, so I have to thank Jagex for getting my school grades up.
The grand exchange (auction house) and the trade restrictions that landed at the same time pretty much killed the game for me.
Prior to the G.E., RuneScape had a thriving, complex economy. Players made money transporting goods from harvesting areas, either on behalf of someone or by paying independents for their goods so the independents could avoid going back to town. Players made money buying and selling goods - geographic arbitrage was very much a thing, as well as across time, and also across servers. People made money turning cheaply available goods into more expensive goods.
When the G.E. landed, it basically killed most of the economy. Harvesting stuff could still be profitable, and players could still make money transporting goods from harvesting areas to the nearest bank so people grinding levels wouldn't have to leave, but basically everything else became irrelevant. There's no point in selling anything anywhere yourself when you could take it to the G.E. and get a sale with no effort.
Less of an issue but still sad, the trading restrictions also killed the generosity of veteran players. 'drop parties', where a rich player leads a group around town dropping valuable items, died off as valuable items would no longer appear in the ground for others. Gifting people stuff was no longer possible if it exceeded (fairly low) thresholds. Very sad.
How is that any different? It would be fine if schools banned children from bringing cash to school. If there were a multibillion dollar bullying and gambling market going on at the school, I'd demand it.
My issue with that is that the kids are losing something because someone else is doing something. Very similar to one kid being disorderly in class and everyone losing the recess.
Imagine being a kid in that room and being annoyed by the kid being disorderly, because you want to learn. Now you lost your rights because of that kid. You never did what he was doing, you never contributed to the disorder he caused, if anything you were also victimized by it. And then the power figure in this equation goes and chops away your rights along with his. First lesson in unfairness where the wet grass is burnt alongside the dry grass, because to the powers that be, the rights and allowances you had are mere acceptable collateral damage. Suppressing dissent was more important than protecting what is yours.
Are these schools in Stalin's soviet union? One kid causes disorder so all kids must be purged to make sure there won't be another naughty child in the future?
Believe it or not, teachers (your sao-called "power figures" here) are generally not a bunch of untrained dumbshits unable to think of kids with more granularity than as the entire collective group making up a class. They have the skills and training to identify the sources of disruptions along with ample resources available for correcting them without calling forth damnation and hellfire on everything in a 5 mile radius. Hammers are awesome, but it's not all that hard to grab a scalpel when a situation calls for a scalpel.
With trading. Market price is not only price. With skins there are lot of preferences. Starting for weapons they apply, not all people use same guns as much. And they might prefer one style of skin over an other skin.
If you are not in to extract most possible value, you might trade a more expensive skin for weapon you do not use in style you do not use for less expensive one for weapon you use more in style you really like.
This sounded odd to me as well. Most lootbox games today has no trading at all - you can pay to unlock items for your own account, and that's it. I suspect the more accurate way to explain it is Valve successfully prevented authority interventions than this being more consumer friendly.
One of the difficulties -- and one that is currently a big problem in LLM research -- is that comparisons with or evaluations of commercial models are very expensive. I co-wrote a paper recently and we spent more than $10,000 on various SOTA commercial models in order to evaluate our research. We could easily (an cheaply) show that we were much better than open-weight models, but we knew that reviewers would ding us if we didn't compare to "the best."
Even aside from the expense (which penalizes universities and smaller labs), I feel it's a bad idea to require academic research to compare itself to opaque commercial offerings. We have very little detail on what's really happening when OpenAI for example does inference. And their technology stack and model can change at any time, and users won't know unless they carefully re-benchmark ($$$) every time you use the model. I feel that academic journals should discourage comparisons to commercial models, unless we have very precise information about the architecture, engineering stack, and training data they use.
In your rush to condemn it, did you actually read the linked article? Let me quote:
"These tests [...] were “supposedly applicable to both white and black prospective voters who couldn’t prove a certain level of education” (typically up to the fifth grade). Yet they were “in actuality disproportionately administered to black voters.”
Additionally, many of the tests were rigged so that registrars could give potential voters an easy or a difficult version, and could score them differently as well. For example, the Veterans of the Civil Rights Movement describes a test administered in Alabama that is so entirely subjective that it measures the registrar’s shrewdness and cunning more than anything else."
The bad news is this "enlightened racism apologist" trend seems to be on the uptick. The good news is that they're so unfathomably ignorant of basic historical facts (as you demonstrate) that it's easy to prove they're actually mere average-intelligence racism apologists.
I'd recommend trying Instant Neural Graphics Primitives (https://github.com/NVlabs/instant-ngp) from NVIDIA. It's a couple years old, so not state-of-the-art, but it runs on just about anything and is extremely fast.
reply