Shape matching plays an important role\nin various computer vision and graphics applications\nsuch as shape retrieval, object detection, image editing,\nimage retrieval, etc. However, detecting shapes in\ncluttered images is still quite challenging due to the\nincomplete edges and changing perspective. In this\npaper, we propose a novel approach that can efficiently\nidentify a queried shape in a cluttered image. The core\nidea is to acquire the transformation from the queried\nshape to the cluttered image by summarising all pointto-\npoint transformations between the queried shape\nand the image. To do so, we adopt a point-based shape\ndescriptor, the pyramid of arc-length descriptor (PAD),\nto identify point pairs between the queried shape and\nthe image having similar local shapes. We further\ncalculate the transformations between the identified\npoint pairs based on PAD. Finally, we summarise\nall transformations in a 4D transformation histogram\nand search for the main cluster. Our method can\nhandle both closed shapes and open curves, and is\nresistant to partial occlusions. Experiments show that\nour method can robustly detect shapes in images in\nthe presence of partial occlusions, fragile edges, and\ncluttered backgrounds.
Loading....