A Novel Self-Intersection Penalty Term for Statistical Body Shape Models and Its Applications in 3D Pose Estimation

Round  1

Reviewer 1 Report

The authors should compare their approach, both quantitatively and qualitatively, to other state of the art methods. Otherwise the scientific relevance of the proposal cannot be assessed.

Author Response

Response to Reviewer 1 Comments

Point 1: The authors should compare their approach, both quantitatively and qualitatively, to other state of the art methods. Otherwise the scientific relevance of the proposal cannot be assessed.

Response 1: Considering this suggestion, we have conducted the quantitative comparison and qualitative comparison with two state-of-the-art methods of avoiding self-intersection of triangular mesh in the field of 3D reconstruction and 3D pose estimation. One of the methods to compare is Laplacian regularization which prevents the vertices from moving too freely and avoids self-intersection potentially. The other is sphere approximation which approximates the body shape with a set of spheres and avoids self-intersection by discouraging the intersection between spheres. Per vertex error and percentage of vertices in self-intersection were employed as our evaluation metrics for quantitative evaluation. According to the experimental results, our proposed method outperforms these two approaches both qualitatively and quantitatively.

Author Response File: Author Response.pdf

Reviewer 2 Report

The algorithm is interesting and the results are good. I would appreciate more clarity in the description.

Figure 1 does not fulfill its purpose because the self intersections are not clearly observed. It is difficult to differentiate the two way of self-intersection in Figure 2.

The paper says that "the density of rays depends on the number of triangles in the mesh" but after the number of rays depends of the pixels in the image. I would appreciate to clarify this point.

I consider that the vertex classification algorithm can classify the same vertex in different group, depending if one of its triangle intersects or not. How is this handled?

Author Response

Response to Reviewer 2 Comments

Point 1: The algorithm is interesting and the results are good. I would appreciate more clarity in the description.

Response 1: Thank you for your favorable comments about our work and we tried our best to revise and improve the description in our manuscript.

Point 2: Figure 1 does not fulfill its purpose because the self intersections are not clearly observed. It is difficult to differentiate the two way of self-intersection in Figure 2.

Response 2: We are sorry for our improper representations in Figure 1 and Figure 2. We picked two examples of body mesh with self-intersection which can be clearly observed to replace the original images in Figure 1. And we added an intuitive representation of the process of how a spherical surface deforms to the two way of self-intersection into Figure 2 to make these images more distinctive.

Point 3: The paper says that "the density of rays depends on the number of triangles in the mesh" but after the number of rays depends of the pixels in the image. I would appreciate to clarify this point.

Response 3: Thank you for your careful reading and pointing out this contradiction in our manuscript. We are sorry that our inappropriate description makes it confusing to understand. We have rewrite this part of description in Line 157 to clarify that the density of detection rays is manually set to maintain the accuracy of classification and low memory consumption according to the number of triangles in the mesh. And the "camera" and "pixel" we mentioned in the manuscript are just used as the source from where the detection rays are emitted. We use these terminologies for the sake of description.

Point 4: I consider that the vertex classification algorithm can classify the same vertex in different group, depending if one of its triangle intersects or not. How is this handled?

Response 4: It is really true as you considered that the same vertex can be classified into different groups. This will happen when the vertex is near to the boundary of the self-intersection region and the neighbouring triangles of the vertex are in different region. To handle this problem when implementing the code, we assume that a vertex is in self-intersection if at least one of neighbouring triangle is in self-intersection or else this vertex is not in self-intersection. This assumption is nearly true when the triangular mesh is dense enough and the triangles are small enough to assume that each triangle and its vertices are located in the same region. But our assumption does not apply to the situation when the vertices in the mesh are sparse and triangles are very large, it is probably that undesirable consequences may be caused if our proposed method is employed in this situation. Moreover, we added this to the section of conclusion as a limitation of our method.

Author Response File: Author Response.pdf

Round  2

Reviewer 1 Report

The reviewers have addressed my previous concerns, so I recommend acceptance.