Title: Its a 3D World, After All
1Its a 3D World, After All
2The sad and miserable life of an object detector
3People Detection in the Middle Ages
The Empress Theodora with her court. Ravenna,
St. Vitale 6th c.
4Multiscale processing
5Many Object Categories
6Occlusion, Pose
Giotto, The Mourning of Christ, c.1305
7Perspective
East Doors (1452)
North Doors (1424)
Lorenzo Ghiberti (1378-1455)
8Perspective gone wild
Piero della Francesca, The Flagellation (c.1469)
9Real World
10Objects vs. Scenes
11Scene Understanding (The Age of Titans)
- Guzman (SEE), 1968
- Hansen Riseman (VISIONS), 1978
- Barrow Tenenbaum 1978
- Brooks (ACRONYM), 1979
- Marr (2 ½ D sketch), 1982
- Ohta Kanade, 1978
12What Went Wrong?(The Age of Titans)
13Learning Geometry
Torralba Oliva, 2001
14Learning Geometry
Hoiem, Efros, Hebert, 2005
Andrew Ng et al, 2006
15Automatic Photo Pop-up
Geometric Labels
Original Image
16The World Behind the Image
Automatic Photo Pop-up, SIGGRAPH05
17Parsing Whole Scene
18Scene
19Objects and Scenes
- Biedermans violations (1981)
20Probability, position (2D)
Torralba et al
21Size
Torralba et al
22Position (3D)
Saddeth, Torralba, Freeman, Wilsky, 2006
23Support, Size
2
?
3
?
1
?
24Improving Object Detection
25Improving Object Detection
26Improving Object Detection
27Improving Object Detection
28Hoiem, Efros, Hebert, 2006
29Qualitative Results
30Top View
31Interposition, Depth Layers
32Discussion Questions
- 1. Can we inject 3D knowledge into
appearance-based methods? - 2. Should we build datasets supporting 3D shape?
- 3. Where does segmentation come in?
- 4. Is 3D scene modeling necessary?
- 5. Is explicit 3D (e.g. top-down view) necessary?
- 6. Is depth layer extraction the right problem?
How to approach it?