Global alignment of deformable objects captured by a single RGB-D camera

Abstract

We present a novel global registration method for deformable objects captured using a single RGB-D camera. Our algorithm allows objects to undergo large non-rigid deformations, and achieves high quality results without constraining the actor's pose or camera motion. We compute the deformations of all the scans simultaneously by optimizing a global alignment problem to avoid the well-known loop closure problem, and use an as-rigid-as-possible constraint to eliminate the shrinkage problem of the deformed model. To attack large scale problems, we design a coarse-to-fine multi-resolution scheme, which also avoids the optimization being trapped into local minima. The proposed method is evaluated on public datasets and real datasets captured by an RGB-D sensor. Experimental results demonstrate that the proposed method obtains better results than the state-of-the-art methods.

Global as-Conformal-as-Possible Non-Rigid Registration of Multi-view Scans

“…We first evaluate our method on Jumping dataset [16] which contains complete models with large deformations and known correspondences, compared with three state-of-theart methods [17,9,10]. For pairwise registration methods [17,9], we register all the models in sequence with the previous registration result used as the next target model.…”

“…Fortunately, the fact is observed and verified that the deformations of usual objects such as human bodies and animals are locally rigid. Therefore, we use an orthogonality constraint similar to [10]:…”

“…where D ∈ R 3×4 is defined to extract the rotation transformation from X . The ARAP constraint is defined as [12,13]:…”

“…Multi-Resolution Approach: Considering the high dimensionality and complexity of the deformation space in nonrigid registration, we use a coarse-to-fine approach [10] to reduce the degrees of freedom and deal with the large scale problems efficiently. Suppose that we decompose the shapes up to S +1 scales.…”

“…Li et al [9] present a pair-wise non-rigid registration method for large deformation cases, but the loop closure problem tends to occur due to error accumulation. Guo et al [10] propose a global non-rigid registration method to avoid the error accumulation, but the viewpoints need to be reasonably dense to avoid too large differences between neighboring views.…”