recovering surface layout from a single image d. hoiem, a.a. efros, m. hebert robotics institute,...

55
Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29, 2009

Upload: juliana-randall

Post on 31-Dec-2015

222 views

Category:

Documents


4 download

TRANSCRIPT

Page 1: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Recovering Surface Layout from a Single Image

D. Hoiem, A.A. Efros, M. HebertRobotics Institute, CMU

Presenter: Derek HoiemCS 598, Spring 2009

Jan 29, 2009

Page 2: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Why worry about 3d scenes?

Page 3: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Reason 1: We may want to interact with the scene

Navigation Manipulation

Page 4: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

4

Reason 2: We need context

Page 5: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

5

Reason 2: We need context

Page 6: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

2D Object Detection

Page 7: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

What the 2D Detector Sees

Page 8: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Computers need context tooTrue

Detection

True Detections

MissedMissed

False Detections

Local Detector: [Dalal-Triggs 2005]

Page 9: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

9

Context in Image Space

[Kumar Hebert 2005][Torralba Murphy Freeman 2004]

[He Zemel Cerreira-Perpiñán 2004]

Page 10: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

We need 3d info to reason about 3d relationships

Close

Not Close

Page 11: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

How to represent scene space?

Page 12: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

How to represent scene space?

Holistic Scene Space: “Gist”

Oliva & Torralba 2001

Torralba & Oliva 2002

Page 13: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

How to represent scene space?

Depth Map

Saxena, Chung & Ng 2005, 2007

Page 14: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Gibson’s Surface Layout

slide from Aude Oliva

• Gibson: “The elementary impressions of a visual world are those of surface and edge.” The Perception of the Visual World (1950)• Focus on texture gradients

Page 15: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Surface Layout (Gibson cont.)

slide from Aude Oliva

Gibson’s Surface Layout

Page 16: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Surface Layout (Gibson cont.)

slide from Aude Oliva

Gibson’s Surface Layout

Page 17: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Marr’s 2½D Sketch

Marr’s 2½-D Sketch

Figs from Aude Oliva slide

Page 18: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Surface Layout (this paper)

Goal: Label image into 7 Geometric Classes:• Support• Vertical

– Planar: facing Left (), Center ( ), Right ()– Non-planar: Solid (X), Porous or wiry (O)

• Sky

Page 19: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Our Main Challenge

• Recovering 3D geometry from single 2D projection

• Infinite number of possible solutions!

Page 20: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Our World is Structured

Abstract World Our World

Image Credit (left): F. Cunin and M.J. Sailor, UCSD

Page 21: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Most Early Work Tried to Manually Specify the Structure

• Hansen & Riseman 1978 (VISIONS)• Barrow & Tenenbaum 1978 (Intrinsic Images)• Brooks 1979 (ACRONYM)• Marr 1982 (2½ D Sketch)

Ohta & Kanade 1978Guzman 1968

Page 22: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Learn the Structure of the World

Page 23: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Infer Most Likely Scene

Unlikely Likely

Page 24: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

1. Use All Available Cues

Vanishing points, lines

Color, texture, image location

Texture gradient

Page 25: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Use All Available Cues

Page 26: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

2. Get Good Spatial Support

50x50 Patch50x50 Patch

Page 27: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Image Segmentation

• Single segmentation won’t work

• Solution: multiple segmentations

Page 28: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

For each segment:

- Get P(good segment | data) P(label | good segment, data)

Labeling Segments

Page 29: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Image Labeling

Labeled Segmentations

Labeled Pixels

segments

datasegmentgoodlabelPdatasegmentgoodPdatalabelP ),|()|()|(

Page 30: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

30

Gray?

High inImage?

Many LongLines?

Yes

No

NoNo

No

Yes Yes

Yes

Very High Vanishing

Point?

High in Image?

Smooth? Green?

Blue?

Yes

No

NoNo

No

Yes Yes

Yes

Decision Trees + AdaboostDecision Trees + Adaboost

Ground Vertical Sky

Collins et al. 2002

Page 31: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Surface Confidence Maps

P(Support) P(Vertical) P(Sky)

P(Planar Left) P(Planar Center) P(Planar Right)

P(Non-Planar Porous) P(Non-Planar Solid)

Test Image

Page 32: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Experiments: Input Image

Page 33: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Experiments: Ground Truth

Page 34: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Experiments: Our Result

Page 35: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Surface Estimates: Outdoor

Input Image Ground Truth Our Result

Avg. Accuracy

Main Class: 88%

Subclass: 62%

Page 36: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Input Image Ground Truth Our Result

Surface Estimates: Outdoor

Page 37: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Input Image Ground Truth Our Result

Surface Estimates: Outdoor

Page 38: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Surface Estimates: Paintings

Input Image Our Result

Page 39: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Surface Estimates: Indoor

Avg. Accuracy

Main Class: 93%

Subclass: 76%

Input Image Ground Truth Our Result

Page 40: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Failures: Reflections and Shadows

Input Image Our Result

Page 41: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Average Accuracy

Main Class: 88%

Subclasses: 61%

Page 42: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Importance of Many Cues

All Position Only

Color Only

Texture Only

Perspective Only

Main 88% 83% 72% 80% 68%

Subclass 61% 43% 43% 55% 52%

All All But Position

All But Color

All But Texture

All But Perspective

Main 88% 84% 87% 87% 88%

Subclass 61% 60% 60% 58% 57%

Page 43: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Importance of Many Cues

Page 44: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Spatial Support Matters

Page 45: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Automatic Photo Popup

Labeled Image Fit Ground-Vertical Boundary with Line

Segments

Form Segments into Polylines

Cut and Fold

Final Pop-up Model

[Hoiem Efros Hebert 2005]

Page 46: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

video

Page 47: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Surfaces Not Enough – Need Occlusion Reasoning

Image Surface Labels 3D Model

Page 48: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Surfaces + Occlusions + Objects = Better 3D Models

Surfaces Occlusions

Objects and Viewpoint

SupportHorizon, Object Maps

Surface Maps

Depth, Boundaries

Boundaries

Horizon, O

bject Maps

Viewpoint/Size Reasoning

Page 49: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

video 2

Page 50: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Contributions• General principles

– Learn the structure of the world– Use all available cues– Spatial support matters– Use redundancy to deal with unreliable processes

(segmentation)

• Results include entire spread of failure and success

• First work to convincingly demonstrate single-view reconstruction

Page 51: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Criticisms• Still just 2D pattern recognition?

• Not clear how to generalize to arbitrary 3d angles

• Restricted to visible portion of scene

• Coarse layout: not clear if applicable to personal space or object shapes

Page 52: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Ideas for improvement

• Try improving features (e.g., add bag of words)

• Extend to characterize object shapes?

• Combine this surface-based layout with depth estimates from Saxena et al.

Page 53: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Discussion• Use for context (Eamon)• Multiple segmentations (Duan, Sanketh)• Subcategories (Duan, Sanketh)• Global info, use of object knowledge (Binbin)• Combination with multiview cues (Mani)• Landmarks (Gang)

Page 54: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Thank you

Page 55: Recovering Surface Layout from a Single Image D. Hoiem, A.A. Efros, M. Hebert Robotics Institute, CMU Presenter: Derek Hoiem CS 598, Spring 2009 Jan 29,

Things to cover when you present

• Background• Overview of method• Results• Things you like• Things you don’t• Ideas for improvement• Address bulletin board postings