Shape matching and object recognition are widely researched areas in 3D computer vision, with applications ranging from reconstruction to surveillance. On the one hand, shape matching concerns the problem of determining a dense correspondence between two given objects. On the other hand, object recognition consists in locating, and at the same time putting into correspondence a template model within a given scene which contains the object of interest.
A particularly challenging instance of this problem arises when the object to be sought is allowed to deform in a non-rigid fashion – a common scenario, for instance, in robotics applications, where one has to locate a reference model within a dynamic environment acquired in 3D.
Despite the conceptual similarities, however, 3D shape matching and recognition have been tackled separately and under different assumptions. Deformable matching techniques assume the absence of additional objects (clutter); conversely, object-in-clutter methods rely on the scene to contain a rigidly transformed instance of the model. These constraints severely limit the usefulness of either family of approaches in practical scenarios.

We rule out all the previous assumptions and consider the full problem of deformable objectin-clutter recognition and matching. For given model and scene, both of which are allowed to deform non-rigidly, we jointly determine the object location in the scene and solve for a dense correspondence between the two.

Figure: Some solutions obtained by our pipeline for deformable object-in-clutter. Corresponding points have same color; white color denotes no match. Observe how the correspondence is accurate also for points close to scene boundary.


Download code and data (5.2GB) zip
The code has been compiled and tested with Matlab 2016b and Microsoft Visual C++ 2012 (C) on Windows 8.1.