1
00:00:12,000 --> 00:00:14,000
I'm Harlyn Baker

2
00:00:14,000 --> 00:00:17,000
I'm a long time 3D vision researcher and I'm the founder of EPIImaging

3
00:00:18,000 --> 00:00:22,799
3D computer vision has been an academic endeavor for many decades and with a

4
00:00:22,800 --> 00:00:27,000
huge number of applications developing now because of mobile computing and

5
00:00:27,179 --> 00:00:30,809
effectively the cell phone wars that have brought us great advances in

6
00:00:30,809 --> 00:00:33,000
cameras and processors, it is becoming very important.

7
00:00:34,000 --> 00:00:38,519
It has applications in augmented reality and virtual reality where you want to

8
00:00:38,520 --> 00:00:43,500
both capture scenes to project to people and to know where people are looking in the world,

9
00:00:43,800 --> 00:00:46,890
what their world looks like, to put objects into their environment

10
00:00:47,820 --> 00:00:51,900
it's important in self-driving vehicles like cars and drones 

11
00:00:52,000 --> 00:00:53,819
where you need to have 3D geometry

12
00:00:53,820 --> 00:00:58,920
it's important in robotics particularly maybe humanoid robotics where we have

13
00:00:58,920 --> 00:01:02,579
to deal with devices that will soon be walking among us and having to interact

14
00:01:02,579 --> 00:01:04,920
with all the uncertainties of the world.

15
00:01:04,920 --> 00:01:10,019
The technology I am promoting is one that we developed a long time ago,

16
00:01:10,019 --> 00:01:13,859
back in the days of single cameras, and we asked the question back then 

17
00:01:13,859 --> 00:01:16,000
what happens, what will happen

18
00:01:16,000 --> 00:01:18,420
when cameras are tiny and when they're cheap

19
00:01:19,200 --> 00:01:22,100
basically free; what happens when the bandwidth that we were 

20
00:01:22,099 --> 00:01:24,500
struggling with is no longer a problem

21
00:01:24,500 --> 00:01:29,320
and we developed methodologies that sort of took care of that, that said 

22
00:01:29,319 --> 00:01:31,029
this is what you might do, and that's what

23
00:01:31,030 --> 00:01:35,718
in fact we're doing now, we're looking at solutions that involve a different approach

24
00:01:35,718 --> 00:01:40,000
and its an approach that is coupled with something called Light Fields

25
00:01:40,000 --> 00:01:44,019
where you take many many cameras, you capture a huge amount of data 

26
00:01:44,019 --> 00:01:47,500
much more than people might think is necessary but that isn't the point

27
00:01:48,250 --> 00:01:53,349
these cameras are free, cameras are tiny, 
and the bandwidth is all doable - 

28
00:01:53,349 --> 00:01:56,530
its all handleable by our new capabilities.

29
00:01:56,530 --> 00:02:01,030
So what can we do with that?  Well, what we 
can do is we can have solutions that

30
00:02:01,030 --> 00:02:04,590
capitalize on redundancy, and 
that give us results that have 

31
00:02:04,590 --> 00:02:09,310
much higher accuracy and much greater precision; 
important features

32
00:02:09,310 --> 00:02:13,449
you don't want systems that make mistakes.  
So with the extra cameras that

33
00:02:13,449 --> 00:02:18,000
we use, and I use tens, twenties, thirties, hundreds, and
I'm building them into fingertips of robots,

34
00:02:18,500 --> 00:02:20,920
I'm building them into drones,

35
00:02:20,919 --> 00:02:25,500
I'm building them into vehicles like cars, 
I'm building them into light-field cameras

36
00:02:26,000 --> 00:02:28,989
for capturing immersive cinema -- about a meter by meter -- all these

37
00:02:28,989 --> 00:02:31,209
are massive numbers of cameras

38
00:02:31,209 --> 00:02:35,439
thousands perhaps, different sizes, different characteristics, but their purpose

39
00:02:35,439 --> 00:02:39,969
is to enable bringing in the constraints that one can have with these analyses 

40
00:02:39,969 --> 00:02:45,250
when you get beyond the old approaches, and use even older approaches

41
00:02:45,250 --> 00:02:49,239
but that were perhaps ahead of their time, that are now viable.

42
00:02:49,780 --> 00:02:53,318
So the challenge is to use these technologies and these advances

43
00:02:53,318 --> 00:02:56,318
to capture the 3D geometry of the world.

44
00:02:57,469 --> 00:03:02,449