A Mid-Level Approach for Efficient Object Search in Clutter

Abstract:

We describe an efficient contour-based object recognition technique suitable for searching objects in clutter. The key innovation is the use of high-level knowledge concerning the target object so that mid-level visual processes are directly influenced. Using the recently introduced image torque operator, potential locations of objects are first detected as attention points. At each fixation point, contours that are supporting the torque are grouped and matched in a hierarchical manner using a novel shape representation that uses the torque information to enforce boundary ownership and rotational invariance. The matching scores are then converted to weights to modulate the torque operator in a second pass so that the target location is immediately highlighted. The approach is validated over a: 1) hand manipulation task involving a robot observing a table with tools and objects in various poses, under clutter and occlusion and 2) shape-based retrieval image dataset of complex scenes.