Fast Census Transformbased Stereo Algorithm using SSE2

1
Fast Census Transform-based Stereo Algorithm
using SSE2

• Young Ki Baik
• Kyoung Mu Lee
• Computer Vision Lab.
• School of Electrical Engineering and Computer
  Science
• Seoul National University

2
Contents

• Stereo Vision
• Census Transform Stereo Vision
• Fast approaches
• Experimental result
• Conclusion and Future work

3
Introduction

• What is the stereo vision?
• The stereo vision is the method to extract 3D
  information using image from different view
  points.
• Topographical survey
• Obstacle detection
• Object tracking
• Face recognition

4
Introduction

• Trade off of algorithms
• Algorithm for accurate results
• Complex computation and iteration
• Slow processing time
• Unable to realize real-time system
• Algorithm for fast processing time
• Simple computation and no iteration
• Fast processing time
• Unable to realize accurate system

5
Introduction

• Fast stereo vision algorithm
• Window size invariant method
• Box filtering method
• Box-filtering techniques, M.J.McDonnell
  (CGIP-81)
• Real time correlation-based stereo algorithm,
  implementations and applications, Olivier
  Faugeras , Zhengyou Zhang , (Tech.Rep.RR-2013,
  INRIA,1993)
• Disparity range invariant method
• Rectangular subregioning method
• Rectangular Subregioning and 3-D Maximum-Surface
  Techniques for Fast Stereo Matching, Changming
  Sun (CVPR-2001)
• Parallel processing technique

6
Introduction

• Problem
• Real images from grabbers can not assure of
  brightness consistency in corresponding region.
• Intensity correlation method is not proper for
  real images.
• Census transform
• Census transform has been evaluated as the method
  robust to radiometric distortion.
• J. Banks and P. Corke, "Quantitative evaluation
  of matching methods and validity measures for
  stereo vision," Int. J. Robotics Research, vol.
  20, pp. 512-532, July 2001.
• Heiko Hirschmller, "Improvements in Real-Time
  Correlation Based Stereo Vision", Proceedings
  IEEE Workshop on Stereo and Multi-Baseline
  Vision, pp. 141-148, Kauai, Hawaii, December 2001.

7
Census Transform Stereo Vision

• Census transform
• Census transform converts relative intensity
  difference to 0 or 1 and deforms 1 dimensional
  vector as much as window size of census
  transform.

8
Census Transform Stereo Vision

• Result of census transform
• Census transform makes data of (image size
  vector size).

(Square size of CTW)-1
Height
Width
9
Census Transform Stereo Vision

• Sum of Hamming distance
• The Hamming distance of two transformed vectors
  with correlation windows is used to find
  corresponding region.

Sum of Hamming distance
Disparity range
Right census transformed vector
3D disparity space
Left census transformed vector
10
Census Transform Stereo Vision

• Complexity of algorithm
• Census transform-based stereo vision (CTSV) has
  high complexity.
• N Searching window size
• D Disparity range
• C Census transform window size

11
Fast Census Transform Stereo Vision

• Fast Census Transform Stereo Vision

Census transform
Hamming distance
Parallel processing - SSE2
8bit look-up table Parallel processing - SSE2
Correlation
Moving window technique Parallel processing - SSE2
12
Fast Census Transform Stereo Vision

• SSE2 (Streaming SIMD Extension 2)
• 128-bit SIMD packed integer floating point
  arithmetic operation
• Cache and memory management operation
• Continuous memory structure is required
• No advantages in separate data

13
Fast Census Transform Stereo Vision

• Fast approaches (census transform)
• Usage of parallel processing (SSE2)
• 16 pixels are loaded to XMM(SSE2 memory) and
  computed at once.

14
Fast Census Transform Stereo Vision

• Fast approaches (sum of Hamming distance)
• Usage of 8bit look-up table (LUT)
• Parallel processing SSE2
• Parallel processing is faster than 8 bit LUT

15
Fast Census Transform Stereo Vision

• Fast approaches (correlation)
• Moving window technique

16
Fast Census Transform Stereo Vision

• Fast approaches (correlation)
• Combination of moving window and SIMD
• Moving window technique for x, y-axis
• SSE2 for d-axis

17
Experimental result

• Environment
• System Pentium-IV 2.4GHz
• Cache memory 512Kbyte
• Camera Stereo Mega-D (Videre design)
• Condition
• Image size 320 x 240 gray stereo images
• Census transform window size 5x5, 7x7, 9x9
• Disparity searching range 32
• Correlation window size 11x11

18
Experimental result

• Detail processing time
• 32 disparity searching range

19
Experimental result

• Performance of census transform stereo vision

20
Conclusion and Future work

• Conclusion
• Moving window technique reduces processing time
  to constant except in transforming stage
• SSE2 instructions reduces running time by 2.5 to
  3 times
• Possibility for faster result
• Specialized Instruction
• 16 bit look-up table
• Fixed window size of census transform
• Future work
• Applying real-time approach to another stereo
  algorithm
• Combine stereo system to other applications