Measuring Stereoscopic Image Compression Quality

1
Measuring Stereoscopic Image Compression Quality
2

• Stereoscopic Compression Methods
• Symmetric Compression
• Asymmetric Compression
• Image and Depth Map Interpolation
• Experiments
• Compression
• Peak Signal to Noise Ratio (PSNR)
• Sum of Absolute Differences (SAD)
• Symmetric and Asymmetric Compression Experiment
• Image and Depth Map Interpolation
  Experiment
• Results and Concluding Remarks

3
Stereoscopic Compression Methods
4
Stereo Compression Methods
Stereo Compression Methods

• Compression plays a significant role in the
  storage and transmission of digital video data
  and images.
• Transmission of two images is required, left
  and right frames
• Twice the Bandwidth is needed

Symmetric Compression

• Symmetric stereo compression applies equal
  amounts of compression to each image, left and
  right, of a stereo pair.

Asymmetric Compression

• Asymmetric stereo compression applies unequal
  amounts of compression to each image of a
  stereo pair.

Image and Depth Map

• A single image is used with an accurate depth
  map. The depth map is used to interpolated the
  other view(s) from the single image.

5
Experiments
6

• JPEG and JPEG 2000 Compression
• Way to save valuable bandwidth
• Effect of compression measured to assess
  image quality
• Techniques applied separately to left and
  right views
• Stereoscopic Image Compression
• Opinion divided
• Symmetric or asymmetric compression
• Compress both left and right images
• Are results from depth map interpolation are
  as good as those produced from single
  images?
• Results
• Symmetric image quality is better
• Greater image quality is preserved when
  compressing symmetrically
• Stereo images produced from depth map
  interpolation are very different from the
  original accurate stereo images

7
PSNR

• Peak Signal to Noise Ratio (PSNR)
• Max signal power Corrupting noise power
• Expressed as a logarithmic decibel (dB)
• Calculated using Mean Square Error (MSE)
• Main Use
• To measure image compression quality
• Equations
• I(x,y) is Original Image
• I(x,y) is Approximated Version
• M and N are Image Dimensions
• MAXi Maximum Image Pixel Value

Higher PSNR Better Compression Scheme
8
SAD
Sum of Absolute Differences Metric (SAD)

• Expressed as a Percentage of the Absolute
  Differences between two Images
• Main Use
• To measure image compression quality

Lower SAD Better Compression Scheme
9
Test Images

• Test Images
• Photorealistic computer generated
• Photograph
• Non-Photorealistic computer generated
• Compressed symmetrically and asymmetrically
• Compressed using Depth Map encoding
• Compression using JPEG and JPEG 2000
• Tools designed and used
• To calculate total image file size, for all
  symmetric and asymmetric compressions
• To compare two images using PSNR

10
Compression

• Binocular Suppression Theory
• Binocular percept dominated
• by high quality component

• Results From Existing Subjective Tests
• Show symmetric encoding results in greater
  image quality

These Two Conflicting Methods were Investigated
Further
11
Results
Symmetric/Asymmetric JPEG and JPEG 2000
Compression Results

• For every file size, both JPEG and JPEG 2000,
  symmetric compression out performs the most
  extreme equivalent asymmetric compression.

• It was also observed that, for a given file
  size, the more asymmetric the image pair, the
  lower the quality of the stereo image in
  comparison to the original.

12
Depth Map
A single image is used with an accurate depth
map. The depth map is used to interpolated the
other view(s) from the single image.
Original Uncompressed Stereo Pair
13
Results
Image and Depth Map Interpolation

• For every file size, JPEG symmetric compression
  out performs the equivalent image and depth map
  compression when using both PSNR and SAD as the
  comparison metrics.

14
Concluding Remarks

• Concluding Remarks
• Quality of compressed methods for stereoscopic
  images reviewed
• Results presented
• Results show symmetric compression better for a
  given File Size
• Results showed that stereo images produced
  from depth map interpolation are very different
  from the original accurate stereo images