Image Co-Skeletonization via Co-Segmentation

Recent advances in the joint processing of a set images have shown its advantages over individual processing. Unlike existing works geared towards co-segmentation or co-localization, this article, we explore new topic: image co-skeletonization, which is defined as skeleton extraction foreground objects an collection. It well known that object skeletonization single natural challenging, because there hardly any prior knowledge available about present image. Therefore, resort to idea hoping commonness exists across semantically similar can be leveraged such knowledge, other problems co-segmentation. Moreover, earlier research has found augmenting process with object’s shape information highly beneficial capturing context. Having made these two observations, propose coupled framework for co-skeletonization and tasks facilitate discovery our through process. While primary goal, might also benefit, turn, from exploiting outputs central seeds framework. As result, both benefit each synergistically. For evaluating results, construct novel benchmark dataset by annotating nearly 1.8 K dividing them into 38 semantic categories. Although proposed essentially weakly supervised method, it employed unsupervised scenarios. Extensive experiments demonstrate method achieves promising results all three