Building detection from single airborne images using Markov Random Fields

1
Building detection from single airborne images
using Markov Random Fields

• Authors Antonis Kartazis Prof. Hichem
  SahliPresenter Dr. Frank Cremer

2
Overview

• Objectives
• Assumptions
• Overview of the method
• Grouping hierarchy hypothesis graph
• Hypothesis generations
• Introducing 3-D evidence
• Hypothesis verification
• Results
• Conclusions

3
Objectives

• Detection of individual buildings from
  airborne/satellite images
• Usage single images (e.g. not stereoscopic)
• Model-based approach (appearance model of
  building)

4
Assumptions
Main Assumptions

• We consider that buildings have uniform height
  generally descriptive of flat roofs on top of
  rectangular solids. Gable (slanted) roofs are
  also taken into account as long as the
  base-to-height ratio of the peak or slant is
  small with respect to the overall height of the
  structure.
• A common property that characterizes all
  considered rooftops is the fact that they are
  composed of pair-wise parallel sides.
• Walls are considered to be vertical and each
  building casts its shadow on a surface that is
  locally flat.

5
Overview of the method
Line segment extraction
Edge detection
Original image
Perceptual Organization
Attributed Hypothesis Graph (HG)
Hypothesis generation module

• 2-D/3-D Observation Field
• MRF-based Significance Field
• Height Field

Graph labeling
Identified rooftops
Hypothesis verification module
6
Grouping Hierarchy Hypothesis graph

• The identification of rooftops is treated as a
  perceptual organization problem, based on the
  grouping of image primitive structures
  corresponding to a set of line segments that
  designate object boundaries.
• The grouping of the line segments into more
  abstract structures follows a hierarchical bottom
  up fashion, in which coherent global structures
  gradually emerge from local features.
• Level 0 (S0) Detected line segments
• Level 1 (S1) Continuous lines
• Level 2 (S2) Junctions
• Level 3 (S3) Potential rooftops (polygons with
  pair-wise parallel segments)

7
Grouping Hierarchy Hypothesis graph
Definition of hypothesis graph
8
Grouping Hierarchy Hypothesis graph
9
Hypothesis generation
10
Hypothesis generation
11
Hypothesis generation
Flowchart of the hypothesis generation process.
12
Introducing 3-D Evidence
13
Introducing 3-D Evidence
14
Hypothesis verification
15
Hypothesis verification
16
Hypothesis verification

• Supporting potentials
• For supporting hypotheses the potential is
  designed as follows Two significant grouping
  hypotheses are consistent with each other
  yielding a negative clique potential. A
  significant hypothesis and a hypothesis with low
  significance in contrast are contradicting and
  hence the assigned clique potential is positive.
  Two hypotheses with low significance can neither
  be interpreted as contradicting nor consistent
  resulting in a neutral clique potential of zero.
• Competing potentials
• The clique potentials for two competing
  hypotheses have the opposite effect. Two
  competing hypotheses both being significant
  constitute an inconsistent interpretation and
  contribute a positive potential term to the
  energy. On the other hand, a significant
  hypothesis and a hypothesis with low significance
  are compatible and result in a negative
  potential.
• Use of Iterated Conditional Modes (ICM)
  algorithm for MAP estimation
• Selection of the rooftop hypotheses whose
  significance value exceeds a threshold equal to
  0.5.

17
Results
Generalized gradient
Original image
Detected line segments
Continuous hypotheses
Potential shadow lines
Potential rooftops
18
Results
19
Results
Identified rooftops
20
Conclusions

• We addressed the problem of building detection
  from a single remotely sensed image, without the
  use of predefined parametric models and 3-D
  information provided by multiple views.
• The majority of the existing methods treat the
  2-D grouping independently from the 3-D
  inference. The hypothesis verification step is
  mainly restricted to the highest level of the
  grouping hierarchy (evaluation of the
  hypothesized rooftops) and is usually performed
  in a deterministic manner.
• In the proposed method, the grouping hierarchy
  is treated as a global entity in the form of an
  attributed hypothesis graph (HG) that conveys
  both 2-D and 3-D contextual information of the
  structures of interest (rooftops).
• The hypothesis verification step is formulated
  as a stochastic optimization process that
  operates on the whole HG, in order to find the
  globally optimal
• configuration for the locally interacting
  grouping hypotheses, providing also an estimate
  of the height of each extracted rooftop.
• The proposed stochastic formulation is flexible
  and can be easily adopted in other applications
  related to 3-D object detection/recognition.