Supplementary Material: Semantic Co-segmentation in Videos

We analyze the proposed tracklet co-selection method based on the setting without knowing any prior knowledge on the Youtube-Objects dataset. We first evaluate the importance of facility location F(A) and unary terms U(A) in the submodular function. We show both the intersection-over-union (overlap) ratio for semantic segmentation and the average precision (AP) for classification in Table 1 under the same threshold (i.e., 0.85 as used in the manuscript). With only the facility location term that measures the object similarity, the results are less accurate caused by noisy tracklets, while the unary term can ensure the quality of selected tracklets, and hence produce better results by combining two terms. In Table 2, we show the average overlap ratio over all categories for semantic segmentation with different thresholds applying on re-ranked tracklets. Since a low threshold may result in selecting more tracklets and including more noisy ones, we also report the average F-measure for object classification. Note that we achieve the best result for both segmentation and classification with the threshold 0.75.