Design and Perceptual Validation of Performance Measures for Salient Object Segmentation Vida Movahedi, James H. Elder Centre for Vision Research York University, Canada Evaluation of Salient Object Segmentation Source: Berkeley Segmentation Dataset 2 Centre for Vision Research, York University Evaluation of Salient Object Segmentation How do we measure success? 3 Centre for Vision Research, York University Existing literature Salient object segmentation [Liu07, Zhang07, Park07, Zhuang09, Achanta09, Pirnog09, …] Evaluation of salient object segmentation algorithms [Ge06,?] Evaluation of segmentation algorithms 4 [Huang95, Zhang96, Martin01, Monteiro06, Goldmann08, Estrada09] Centre for Vision Research, York University Contributions Analysis of previously suggested measures Contour Mapping Measure (CM) Order-preserving matching A new dataset of salient objects (SOD) Psychophysics experiments Evaluation of above measures Matching paradigm in Precision and Recall measures 5 Centre for Vision Research, York University Evaluation measures in literature Region-based error measures Based on false positive/ false negative pixels [Young05], [Ge06], [Goldmann08], ... Boundary-based error measures Based on distance between boundaries [Huttenlocher93], [Monteiro06], ... Mixed measures 6 Based on distance of misclassified pixels to the boundaries [Young05], [Monteiro06], ... Centre for Vision Research, York University Region-based error measures [Young05], [Ge06], [Goldmann08], ... A and B two boundaries RA the region corresponding to a boundary A and |RA| the area of this region, False Positives RI ( A, B) 1 RA RB RA RB | RA | RA RB RA RB False Negatives | RB | RA RB RA RB Not sensitive to shape differences 7 Centre for Vision Research, York University Boundary-based error measures [Huang95],[Huttenlocher93], [Monteiro06], ... A and B two boundaries Distance of one point a on A from B is d B (a ) min d (a, b) Hausdorff distance: HD ( A, B) max max d B (a ), max d A (b) Mean distance: MD ( A, B) mean mean d B (a ), mean d A (b) a 8 bB aA aA bB bB Not sensitive to shape differences Centre for Vision Research, York University Mixture error measures [Young05], [Monteiro06], ... Penalizing the over-detected and under-detected regions by their distances to intersection False Negatives 1 MM ( A, B) 2 Ddiag False Positives 1 N fn 1 dA( pj ) N N fp fn j 1 d B ( qk ) k 1 N fp Not sensitive to shape difference 9 Centre for Vision Research, York University Another example Different shapes with low errors 10 Centre for Vision Research, York University Comparing two boundaries B A Small False Negative Region Small False Positive Region The two boundaries need to follow each other Thus it is not sufficient to map points to the closest point on the other boundary The ordering of mapped points must be preserved 11 Centre for Vision Research, York University Order-preserving Mapping The order of mapped points on the two boundaries must be monotonically non-decreasing. If ai bm , a j bn and i j then m n Allowing for different levels of detail: 12 One-to-one Many-to-one One-to-many Centre for Vision Research, York University Contour Mapping Measure Given two contours A=a1a2..an and B=b1b2..bm, Find the correct order-preserving mapping Contour mapping error measure: Average distance between matched pairs of points Bimorphism [Tagare02] Elastic Matching [Geiger95, Basri98, Sebastian03, ..] 13 Centre for Vision Research, York University Contour Mapping Measure A dynamic programming implementation to find the optimum mapping Closed contours point indices are assigned cyclically Based on string correction techniques [Maes90] Complexity: O(nm log m) if m<n and m, n points on two boundaries 14 Centre for Vision Research, York University Contour Mapping Example Ground Truth Boundary Algorithm Boundary Matched pairs shown as line segments CM= average length of line segments connecting matched pairs 15 Centre for Vision Research, York University Contour Mapping Measure Order- preserving mapping avoids problems experienced by other measures 16 Centre for Vision Research, York University SOD: Salient Object Dataset A dataset of salient objects Based on Berkeley Segmentation Dataset (BSD) [Martin01] 300 images 7 subjects 1 1 1 1 1 Source: Berkeley Segmentation Dataset 17 Centre for Vision Research, York University Available in SOD Psychophysical experiments Which error measure is closer to human judgements of shape similarity? 9 subjects 5 error measures 18 Regional Intersection (RI) Mean distance (MD) Hausdorff distance (HD) Mixed distance (MM) Contour Mapping (CM) Centre for Vision Research, York University Psychophysical Experiments Experiment 1 - SOD Reference & test shapes all from SOD Experiment 2 - ALG Reference from SOD, test shapes algorithm-generated Reference: Human segmentation Reference: Human segmentation Test cases: Human segmentations 19 Test cases: Algorithmgenerated Centre for Vision Research, York University Agreement with Human Subjects Human subject chooses Left or Right An error measure M also chooses Left or Right, based on their error w.r.t. the reference shape If M chooses the same as the human, it is a case of agreement Human-Human consistency: defined based on agreement between human subjects 20 Reference Left Centre for Vision Research, York University Right Psychophysical Experiments Experiment 1- SOD Reference & tests shapes all from SOD Experiment 2 - ALG Reference from SOD, test shapes algorithm-generated RI: region intersection, MD: mean distance, HD: Hausdorff distance, MM: mixed measure, CM: contour mapping 21 Centre for Vision Research, York University Precision and Recall measures For algorithm boundary A and ground truth boundary B Precision: proportion of true positives on A Recall: Martin’s PR (M-PR)[Martin04] Minimum cost bipartite matching, cost proportional to distance Estrada’s PR (E-PR)[Estrada09] proportion of detected points on B matched(A, B) | A| matched(B, A) |B| ‘No intervening contours’ and ‘Same side’ constraints Contour Mapping PR (CM-PR) 22 Order-preserving mapping Centre for Vision Research, York University Matching paradigm in Precision/Recall Experiment 1- SOD Reference & test shapes all from SOD 23 Experiment 2 - ALG Reference from SOD, test shapes algorithm-generated Centre for Vision Research, York University Summary Analysis of available measures for evaluation of salient object segmentation algorithms A new measure- contour mapping measure (CM) A new dataset of salient objects Dataset available online: http://elderlab.yorku.ca/SOD Psychophysical Experiment Code available online: http://elderlab.yorku.ca/ContourMapping CM has a higher agreement with human subjects Order-preserving matching paradigm in Precision/Recall analysis 24 Code available online: http://elderlab.yorku.ca/ContourMapping Centre for Vision Research, York University Thank You! 25 Centre for Vision Research, York University