Showing Revision 1 created 11/28/2012 by Amara Bot.

1. Title:
   17-33 Improving Stereo Vision
2. Description:

3. 0 0:00 - 0:06
   I'd like to say a few words about how to improve the results of stereo vision.
4. 6000 0:06 - 0:11
   Here is a vision assembly that James David built up of two cameras.
5. 11000 0:11 - 0:14
   In addition to having these two cameras, he also put a projector into the scene
6. 14000 0:14 - 0:17
   that emitted a random light pattern.
7. 17000 0:17 - 0:23
   In fact, it emitted a striped pattern, shown over here on this frog,
8. 23000 0:23 - 0:29
   and by adding texture to the scene, you can making correspondence easier.
9. 29000 0:29 - 0:33
   This is a striped pattern of unequal distances.
10. 33000 0:33 - 0:38
    There's a coding over here, which makes certain stripes larger than others.
11. 38000 0:38 - 0:41
    If you run the same algorithm I just told you,
12. 41000 0:41 - 0:44
    you'll find that stereo vision becomes better,
13. 44000 0:44 - 0:48
    because we can now better disambiguate the correspondence of points.
14. 48000 0:48 - 0:53
    Here is the assembly used for imaging myself. This is me with a sweater on.
15. 53000 0:53 - 0:55
    That's my face.
16. 55000 0:55 - 1:00
    And you can see by emitting structured light, as it is called,
17. 60000 1:00 - 1:03
    you can enhance the performance of stereo
18. 63000 1:03 - 1:06
    and objects that otherwise have very poor texture.
19. 66000 1:06 - 1:11
    Another solution is called the Microsoft Kinect. You're probably familiar with it.
20. 71000 1:11 - 1:15
    It's a new gaming platform that's been sold at record pace.
21. 75000 1:15 - 1:18
    It uses a camera system, together with a laser.
22. 78000 1:18 - 1:21
    The laser adds texture to the scene,
23. 81000 1:21 - 1:24
    and by triangulation using the same method I showed you,
24. 84000 1:24 - 1:26
    it can recover depth.
25. 86000 1:26 - 1:31
    Here's my postdoc Christian using a Kinect-like sensor
26. 91000 1:31 - 1:36
    to do certain poses in front of a depth sensor.
27. 96000 1:36 - 1:41
    You can see in the screen how his pose is being perceived,
28. 101000 1:41 - 1:51
    and you can see Christian trying to do handstands and other acrobatic maneuvers.
29. 111000 1:51 - 1:54
    He's actually pretty good.
30. 114000 1:54 - 2:07
    That's all using effectively stereo vision.
31. 127000 2:07 - 2:10
    There is actually a whole bunch of different types of techniques
32. 130000 2:10 - 2:13
    for sensing range in computer vision.
33. 133000 2:13 - 2:15
    I'm just going to briefly talk about them.
34. 135000 2:15 - 2:17
    They're called laser range finders.
35. 137000 2:17 - 2:19
    They send off beams of light,
36. 139000 2:19 - 2:22
    and they measure the time until the light comes back into the sensor.
37. 142000 2:22 - 2:25
    They're being manufactured by many different companies.
38. 145000 2:25 - 2:30
    In our experiments using robots to drive through the desert and through traffic.
39. 150000 2:30 - 2:35
    We quite extensively used laser range finders as an alternative to stereo vision,
40. 155000 2:35 - 2:38
    because they give us very, very good range estimates.
41. 158000 2:38 - 2:45
    Here is a 3D model constructed by laser range finders of our neighborhood in Palo Alto,
42. 165000 2:45 - 2:52
    and it's easy to see how 3D points can making amazing 3D models,
43. 172000 2:52 -
    using techniques like stereo vision or like the laser range finders I just briefly talked about.