Paper: 711
Title: A Variational Approach for 3D Shapes Registration by Matching Vector Distance Functions
Abstract:
This paper introduces a new method for 3D shapes registration
by
matching vector distance functions. The vector distance
function
representation is more flexible than the conventional
signed
distance one since it enables us to better control the
shapes
registration process by using more complicated transforms.
Based on
this model, a variational frame work is proposed for the
rigid and
non-rigid registration of shapes which does not need any
point
correspondences. The optimization criterion can handle
efficiently
both the rigid and the non-rigid operations together.
Results in
3D are discussed to show the efficiency of the approach
with small
and large global deformations.
Relevance Score: Novelty Score: Clarity Score: Reference to Prior Work Score: Technical Correctness Score: Experimental Validation Score: Key Contribution: Relevance: Novelty: Reference to prior work: Clarity: Technical Correctness: Experimental Validation : Comments for Author:
Relevance Score: 1 [Of broad interest] Novelty Score: 5 [It is an interesting application] Clarity Score: 2 [Generally clear, though some points require clarification] Reference to Prior Work Score: 2 [References adequate] Technical Correctness Score: 4 [Contains minor errors] Experimental Validation Score: 3 [Quite limited evaluation, but enough to suggest it works] Key Contribution: Shape registration using vector distance functions. Relevance: Novelty: Nothing really new. The only novelty consists in updating common registration approaches to the case of vector distance functions. Reference to prior work: Clarity: Some language mistakes. Line 037: sentence not grammatically correct. Line 171: syntax problem. Line 483: I did not understand it. Technical Correctness: Rotation parameters theta_x, theta_y and theta_z are not introduced; unluckily there is no canonical way to define a rotation with three angles in a cartesian coordinate system. In particular what is a 60,60,60 rotation ? Is a 90,90,90 rotation the Identity ? --- Each time, A is written instead of A(X). --- Line 253: the approximation of the Dirac peaks does not guarantee its integral to be one (which is an important property of Dirac peaks). --- The solution of minimization problem (5) is directly computable without a variational approach .. --- Inhomogeneous scalings, etc, are possible in the usual case of the level-set method: it just requires not to work only with rigid transformations of a fixed level-set but to set the criteria in terms of the shape itself.. --- Since the size of the example images is not given, the size of the translations is not significant. --- The proposed energy claimed to be convex is NOT convex for more complex examples. For instance consider an object which is almost invariant with a 45 degrees rotation, the teeth presented here being almost invariant with a 90 degrees rotation. There will be a local minima for a rotation of 45 degrees... --- Nevertheless, the calculi presented here are perfectly correct. Experimental Validation : The examples converge. This is not really surprising in fact, and not "better" than the ones with usual technics, since they lead to approximatively the same algorithms. Comments for Author: Maybe with the future work you mention at the end of the article, this one could be greatly improved, since by now there is nothing really new or improving results.
Relevance Score: 2 [Of limited but sufficient interest] Novelty Score: 3 [Novel, but mostly incremental] Clarity Score: 3 [Important points are lost because the explanation is not sufficiently clear] Reference to Prior Work Score: 3 [References missing] Technical Correctness Score: 2 [Probably correct (convinced but had to skip some steps)] Experimental Validation Score: 3 [Quite limited evaluation, but enough to suggest it works] Key Contribution: The paper presents a method for 3D shape registration. Shapes are represented by 3- vector distance functions. Relevance: Novelty: The proposed method slightly modifies the work presented in ICCV 2005 (“A shape-based Segmentation approach: An Improved Technique using Level-sets”) – reference [10] in this manuscript. Beyond the slight change of the formulation and the quite limited extension adding penalty term for small deformations, the main contributions of the proposed manuscript: i.e. representation of shapes by vector distance functions (eq. in line 171) and registration model (eq. in line 188) are similar to those presented in [10]. Reference to prior work: The following papers have strong relavance to the proposed work. Chen et. al. “Using prior shapes in geometric active contours in variational framework”. IJCV, 50 (3) 315-328, Dec 2002 Cremers, Kohlberger and Schnorr. “Shape statistics in kernel space for variational image segmentation”. Pattern Recognition, 36(9): 1929-1943, 2003 Cremers and Soatto. “A pseudo distance for shape priors in level set segmentation”. In VLSM pages 169-176, 2003 Riklin-Raviv, Kiryati, Sochen. “Unlevel-sets: geometry and prior-based segmentation”. ECCV , pages 50-61, 2004 Clarity: The text is fairly written and so are the equations. However, the overall organization could be improved, for example Section 5. does not seem to be in the right place. I also failed to understand and thus to evaluate registration results presented in figures 2-3,6 due to insufficient explanations both in text and in the image captions. Specifically, the implication of the red-blue figures is not clear. Technical Correctness: Experimental Validation : I would expect to see the parameters of the recovered transformation with respect to the ground-truth. See my remarks in the Clarity section. Comments for Author: Similar to previous work of the authors, single shape is represented by 3 distance functions. This (awkward) representation is justified, saying that it enables accommodating for a wider group of transformations comparing to the conventional signed-distance function shape representation. Nevertheless, recent works, partially quoted in the ‘Reference to prior work' section present different and possibly simpler shape representations that can accommodate wider group of transformations. The author should refer this issue.
CVIP Lab 