direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Inhalt des Dokuments

Supplementary material for paper:

"Motion-based object segmentation using hysteresis and bidirectional inter-frame change detection in sequences with moving camera"

Marina Georgia Arvanitidou, Michael Tok, Alexander Glantz, Andreas Krutz and Thomas Sikora


Abstract

We present an unsupervised motion-based object segmentation algorithm for video sequences with moving camera, employing bidirectional inter-frame change detection. For every frame, two error frames are generated using motion compensation. They are combined and a segmentation algorithm based on thresholding is applied. We employ a simple and effective error fusion scheme and consider spatial error localization in the thresholding step. We find the optimal weights for the weighted mean thresholding algorithm that enables unsupervised robust moving object segmentation. Further, a post processing step for improving the temporal consistency of the segmentation masks is incorporated and thus we achieve improved performance compared to previously proposed methods. The experimental evaluation and comparison with reference methods demonstrates the validity of the proposed method.

Experimental Evaluation

To demonstrate the segmentation efficiency of our proposed algorithm, we show below the following:

  1. F-Measure (F), Precision (P) and Recall (R) per frame rates (figures) comparing Algorithms 1, 2 and 3
  2. Original, and segmented  video sequences using Algorithms 1, 2 and 3 
  3. P, R and F per frame rates (figures) and average (figures) for different QPs using our proposed Algorithm 3.
  4. Segmented video sequences using Algorithm 3 for QP=38.

Figures may be enlarged upon clicking on the down-right corner of each one.

The original .yuv videos have been compressed to .mp4 files and the segmentated .yuv videos are compressed to .zip files.

.yuv videos can be downloaded and then viewed with an appropriate player such as GLYUVPLAY for mac or YUVPLAYER for windows.

 

Allstars (352 x 288, 250 frames)

Lupe

Nach oben

Biathlon (352 x 288, 200 frames)

What is alternative text oeo?
Lupe

Nach oben

Mountain (352 x 192, 100 frames)

Lupe

Results
Mountain Sequence
Original
Algorithm 1
Algorithm 2
Algorithm 3

Nach oben

Race (544 x 336, 100 frames)

Lupe

Nach oben

Stefan (352 x 240, 300 frames)

Lupe

Nach oben

BBC fish (720 x 576, 120 frames)

Lupe

Nach oben

Horse (352 x 288, 120 frames)

Lupe

Results
Horse Sequence
Original
Algorithm 1
Algorithm 2
Algorithm 3

Nach oben

Results on decompressed data with different bit rates, using the proposed Algorithm 3.

Average Precision per sequence for different quantization parameters.
Lupe
Average F-Measure per sequence for different quantization parameters.
Lupe
Average Recall per sequence for different quantization parameters.
Lupe
Mountain, F-Measure for various QPs
Lupe
Stefan, F-Measure for various QPs
Lupe
Race, F-Measure for various QPs
Lupe
Biathlon, F-Measure for various QPs
Lupe
Allstars, F-Measure for various QPs
Lupe
BBC fish, F-Measure for various QPs
Lupe
Horse, F-Measure for various QPs
Lupe

H.264/AVC settings
KTA 2.4 reference software
IPPP ... GOP structure
EPZS motion estimation
32 x 32 search range
4 x 4 smallest block size



Segmentation results when coded wth QP=38
Mountain, Algorithm4, QP=38
Stefan, Algorithm 4, QP=38
Race, Algorithm, 4 QP=38
Biathlon, Algorithm, 4 QP=38
Allstars, Algorithm, 4 QP=38
BBC fish, Algorithm 4, QP=38  

Horse, Algorithm 4, QP=38
   

Nach oben

Zusatzinformationen / Extras