SegmentationEvaluation

Quantitative evaluation of segmentation results

Evaluation of the multipartition algorithm

• database: Berkeley Segmentation Database
• number of images: 300
• average number of segmentations per image: 5
• explaination of the metrics used:
  • Partition Distances Cardoso et al. http://www.citeulike.org/user/cardelino/article/5043215
  • acronyms:
    • MPD: mutual Partition Distance:
    • APD: asymmetric Partition Distance:
    • SPD: (symetric) Partition Distance:

• URL with the evaluation results: http://iie.fing.edu.uy/rosaluna/juanc/demos/segmentation_evaluation/set_rgb/html/res.html

Misc

• http://www.cs.cmu.edu/~hebert/segs.htm
• http://www1.cs.columbia.edu/CAVE/software/edge/home.php

Datasets

• MSRC
• MSRC2: efros
• berkeley:
  • url: http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/segbench/
  • code:
    • segapp: java tool for drawing segmentations (corrected by me)
• label me:
  • url: http://labelme.csail.mit.edu/
  • user: drunken_sapo
  • todo: download code and see if it can be filtered by user.
  • code:
• PASCAL:
• USF:
  • desc: range images with ground truth
  • url: http://marathon.csee.usf.edu/range/seg-comp/images.html
• image parsing:
  • site: http://www.imageparsing.com
  • mail: "image parsing dataset license" from maggie
  • code: LHI.zip

Papers proposing evaluation methodologies

• cardoso:
  • code:
  • url:
  • ref:
• berkeley:
  • code: cardoso's code computes this on regions (not over boundaries)
  • url:
  • ref:

algorithms that use quantitative evaluation

• http://www.eecs.berkeley.edu/~yang/software/lossy_segmentation/
• no supervisados:
  • calderero (cardoso)
• supervisados (F-measure):
  • arbelaez
  • martin
• segundo plano:
  • TEXTURE-BASED INFRARED IMAGE SEGMENTATION BY COMBINED MERGING AND PARTITIONING
  • Regularized Data Fusion Improves Image Segmentation

Review on evaluation metrics

questions

• Fig. 4 on cardoso's paper. It is not clear which is the test he is running. to get that value of the mutual information.

region based

• martin
  • BCE

• • LCE
    • tolerates mutual refinements
    • $E(S_1,S_2,p_i)=\frac{|R(S_1,p_i) \setminus R(S_2,p_i)|}{|R(S_1,p_i)|}$
  • GCE
    • GCE>LCE
• cardoso
  • SPD
  • APD
  • MPD