31. Step 3 in a bit more detail
• Don’t know what the 3D model should look like
• Don’t know where the photos were taken from
• Do know what matches between different photos
Hey computer, can
you figure this out
for me?
No
problem!
37. Go outside and take
pictures of buildings
Add pictures to flags
Computer
vision
algorithms
Capture flags for your
team and conquer
buildings
Look for flags on the map
38.
39.
40.
41. Feedback for an accepted photo
1058 new points!
Players only earn points for taking useful photos
50. What kind of math went into this?
• Geometry and trigonometry
– How a camera captures a 3D scene in a 2D photos
– How to extract 3D shape from lots of different 2D
photos
– Lots of matrices and linear algebra behind the scenes
(the computer does the hard part)
51. Programming and computers can help
us do more math
• Thousands of
photos
• Millions of 3D
points
• Set up the problem
and have the
computer do the
hard work
Here’s a 2d videogame… mario. Every object has a Width, height… and they move, but they’re FLAT like pieces of paper. always at the same depth on the screen
How many of you have seen minecraft? This minecraft world is made out of things with volume, like lego blocks… things that we can stack
How do you judge depth? Supposetheres a cat in front of us and we want to figure out the distnace to the catUse geometry to figure this out. Make a triangle…
We use geometry to figure out how far away stuff is
Use trigonometry to get this length… kknow the cat is five feet away from the original position.
Don’t have to move around… have two eyesMakes a triangle between our two eyes and some point in the world
When you model this with geometry… triangles
When this camera takes a picture, it looks like this… two pictures from different viewpoints. Slightly differennt…
Toy I grew up with
this poppy device is like the modern version of this, excpet you put in your eyephone
Doesn’t work because the building is in a 3D world, it’s not flat, and the pictures don’t line up
Things that are easy for a computer to detect… really good at corners and where the color changes a lot
We use geometry to figure out how far away stuff is
We use geometry to figure out how far away stuff is
Super technical, but I want to share it with you guys anyway.Gives you shape by moving
Geometry program on the computer figures it out
For us, when we got people to play photocity, the latter happened. And what’s surprising about our game is that while less than 50 people played, and only a small fraction of those made a significant contribution, we still collected over 100,000 photos and put together one of the largest, most comprehensive set of photos of a single location.
This is the website for PhotoCity. It’s got a big map of the campus, and you can click on anything on the map to get more info, which shows up on the side. At the bottom, there’s a bar that says how many flags each team has.
Here’s a more detailed picture of the map with castles and flags representing 3D models in the world that players can grow by taking photos.
Now let me walk you through a cycle of play.The player starts out by identifying a target building on the map…And then they go OUTSIDE and take photosAnd upload them to the flag that they’re targetingWe run our computer vision to find out how much new geometry was created and whether the player was able to capture any flags or conquer any buildings!
Flags are big virtual columns at GPS locations in the world. We’ve colored all the points in this particular flag column orange to direct the players’ attention to expanding this model to the left
When the model grows, the flag gets filled in, and eventually captured by a player.
And then Our system automatically places a new flag as the model grows
Players extend buildings and capture flags by taking and uploading photos. Their photos are automatically accepted or rejected by the system, and when they’re accepted, they actually add 3D points to the 3D model. This shows where the new points came from – they’re the glowing orange points – so players can learn what works about their pictures when they work.
One thing players can do instead of expanding existing buildings, is go out and start their own from scratch. You need anywhere from 20 to 200 photos to start a seed.5 continents around the world
A bunch of different seed models by players
So, now I’m going to tell you about the big competition we had last spring between university of washington and cornell.
Two round that each lasted 3 weeks. We got about 45 players.
This shows the coverage of photos over UW campus – each blue dot is where a photo was taken
And here’s a flythrough of one part of the university of washington campus – this is red square at UW and this model was made from a 3,000-photo subset of all the photos.
Here’s the last video I’m going to show you of the results produced by these players and this game. This particular model was made from 7000 photos and contains almost 20 million 3D points. As computer vision algorithms get even better, we’ll be able to use the original 50,000 photos of uw and 50,000 photos of Cornell and run these through new algorithms and techniques to get even better-looking models.
Cornell arts quad. 5000 images (1/10th of the total images at cornell)As computer vision techniques get better, we’ll be able to rerun these photos and generate clearer models
We’re not doing repetitive math problems by hand
How many of you have an xbox with a kinect? Lets your xbox see in 3D
Right now, kinect is probably the only depth camera you have… you’re going to grow up in a world where these cameras are everywhere… in your cell phone, in your parents’ car
Free program from microsoft that lets you build models from photos
Whenyoureinteractign with it, press c… shows you the geometry of the situation
ive always loved math… math degree in college… ive found that being good at math and programming and using those skills together lets me work on really interesting problems that also involve my hobbies like photography.
Here are some links for you to follow up with at home… photosynth in particular because it’ll do the 3d reconstruction stuff I was talking about for you with your own photos