1. What are the contributions mentioned in the paper "Classifier based graph construction for video segmentation" ?
This is the focus of their paper.. The authors propose to combine features by means of a classifier, use calibrated classifier outputs as edge weights and define the graph topology by edge selection.. By learning the graph ( without changes to the graph partitioning method ), the authors improve the results of the best performing video segmentation algorithm by 6 % on the challenging VSB100 benchmark, while reducing its runtime by 55 %, as the learnt graph is much sparser.
read more
![Figure 1: Climbing up! We consider state-of-the-art features for the computation of superpixel similarities and the graph partitioning theory of recent work [19]. We contribute theory and bestpractices for graph construction and set a new state-of-the-art performance on the challenging VSB100 [20] (BPR and VPR reported here, more details in Section 4.)](/figures/figure-1-climbing-up-we-consider-state-of-the-art-features-hgxmjrdh.webp)
![Figure 5: Comparison of the proposed graph learning method with the baseline algorithm of [19], on the validation set of VSB100 [20]. The plots and table show BPR and VPR measures, aggregate performances ODS, OSS and AP, and length statistics (mean µ, std. δ, no. clusters NCL) (cf. Sec. 3.4 for details).](/figures/figure-5-comparison-of-the-proposed-graph-learning-method-n5s2bzlg.webp)
![Table 2: General applicability of the proposed graph construction. We have tested different clustering methods with the graph of [19] and our learnt graph. In all cases the learnt graph yields better performance and thus generalizes beyond the employed spectral clustering.](/figures/table-2-general-applicability-of-the-proposed-graph-ptnq6yud.webp)
![Figure 6: Comparison of state-of-the-art video segmentation algorithms with our proposed method on the test set of VSB100 [20] (cf. Sec. 4 for details).](/figures/figure-6-comparison-of-state-of-the-art-video-segmentation-2snuzwj9.webp)
![Figure 8: Failure cases for the algorithms [21, 18, 20, 19] and the proposed graph learning method [L(G)]. All methods fail to correctly discern objects, oversegmenting the foreground and background due to the misleading appearance differences and textured background.](/figures/figure-8-failure-cases-for-the-algorithms-21-18-20-19-and-1xoej1e9.webp)