Figure 1. Comparison of GGN with different baselines. (a) an input image from ADE20K [61] (upper row) with the ground truth
object masks overlaid (lower row). Three different approaches: (b) Selective Search (SS) [51], (c) Mask R-CNN [22], and (d) our Generic
Grouping Network (GGN) are applied on the same image to predict the top 100 proposals. The upper images provide all top-100 proposals
predicted by three approaches with false and true positive proposals visualized in red and green boxes, respectively. The lower images
provide only true positive proposals. The number of true positive proposals or ground truth objects are denoted in parentheses. Among 20
ground truth objects, SS recalls only 4, Mask R-CNN detects 5, and GGN retrieves 14. SS is a bottom-up non-parametric approach, thus
has no notion of objectness. Mask R-CNN can make whole object proposals; however it still fails to detect objects that are not seen during
training. Our GGN can predict whole object proposals and generalize to unseen categories.