Categories
Research

Implementing ARAP Deformations using RXMesh

Fellows: Sachin Kishan

Teaching Assistant: Supriya Gadi Patil

Mentor: Ahmed Mahmoud

Introduction

RXMesh is a framework for processing triangle mesh on the GPU. It allows developers to easily use the GPU’s massive parallelism to perform common geometry processing tasks. For a preliminary introduction to RXMesh, check out this tutorial blog here.

Our task was to use RXMesh to implement some classic algorithms in geometry processing. This blog details one of the algorithms we implemented: As-Rigid-As-Possible(ARAP) Surface Modeling [1].

This blog will discuss the algorithm in terms of its basic concepts, the steps of the algorithm, the implementation using RXMesh, and the final results.

Our entire implementation of ARAP deformation in RXMesh can be found here.

As-Rigid-As-Possible Surface Modeling

Sorkine and Alexa (2007) formulated a method where we can retain the shape of an object as much as possible despite deforming one (or a set of) its vertices. This retention of shape is achieved by ensuring that subsequent transformations to the mesh after deforming its vertices are rigid. This property of rigidity means that a mesh can only undergo rigid transformations such as translation or rotation, but not stretching or shearing (since these are not rigid transformations).

The algorithm’s focus is to minimize the energy which measures the extent of deformation (dubbed rigidity energy). This energy is minimized when given a specific set of positions or a specific set of rotations. Notice that we can not minimize both at once—since we need the minimized rotational transformation for minimizing the energy from deformed positions. Thus, we instead solve both alternatively, i.e., solve for deformed positions assuming rotation is fixed, then solve for rotation assuming the position is fixed.

Using a Laplacian, we can form a sparse system of linear equations which we can then solve to obtain a new set of positions for the mesh. The algorithm works as follows for a single iteration

  1. Identify the rotation matrix for the given deformed positions which minimizes deformation energy.
  2. Solve for the new positions to minimize the energy, given a set of linear equations.

In this case, the set of linear equations to solve is in the form :

\(Lx=b\)

Where \(L\) is the Laplace-Beltrami operator, which can be formed by finding the cotangent edge weights of the mesh, \(x\) represents the set of new vertex positions we solve for, and \(b\) represents the right-hand side of Equation 8 from Sorkine and Alexa (2007). The derivation is detailed further in the original paper.

\(b = \sum_{j\in N(i)}^{} \frac{w_{ij}}{2} (R_{i}+R_{j})(p_{i}-p_{j})\)

To further minimize the energy, we can perform multiple iterations of the same steps, using updated vertex positions and consequent rotations per iteration.

Implementation

The following code snippets show how the algorithm is implemented in RXMesh.

We first deform the input mesh and set constraints on each vertex depending on whether it is fixed, deformed, or free to move. This will be taken into account when calculating the system matrix \(L\).

 // set constraints
    const vec3<float> sphere_center(0.1818329, -0.99023, 0.325066);
    rx.for_each_vertex(DEVICE, [=] __device__(const VertexHandle& vh) {
        const vec3<float> p(deformed_vertex_pos(vh, 0),
                            deformed_vertex_pos(vh, 1),
                            deformed_vertex_pos(vh, 2));

        // fix the bottom
        if (p[2] < -0.63) {
            constraints(vh) = 2;
        }

        // move the jaw
        if (glm::distance(p, sphere_center) < 0.1) {
            constraints(vh) = 1;
        }
    });

In the code above, we set the mesh to deform the dragon’s jaw and fix its bottom vertices (legs). This can vary depending on the type of input we want to have. Another possibility is to add a feature for users to click and select vertices to deform or fix them in space interactively.

Before we perform the iteration step, we calculate the Laplacian. It can be calculated as the cotangent edge weight matrix. So, we create a kernel using RXMesh.

    auto calc_mat = [&](VertexHandle v_id, VertexIterator& vv) {
        if (constraints(v_id, 0) == 0) {
            for (int i = 0; i < vv.size(); i++) {
          laplace_mat(v_id, v_id) += weight_matrix(v_id, vv[i]);
          laplace_mat(v_id, vv[i]) -= weight_matrix(v_id, vv[i]);
            }
        } else {
            for (int i = 0; i < vv.size(); i++) {
                laplace_mat(v_id, vv[i]) = 0;
            }
            laplace_mat(v_id, v_id) = 1;
        }
    };

The weight matrix is calculated using a special query called “EVDiamond”, where we obtain the two end vertices and the two opposite vertices of an edge. This gives us exactly the data we need to find the cotangent edge weight.

    auto calc_weights = [&](EdgeHandle edge_id, VertexIterator& vv) {
        // the edge goes from p-r while the q and s are the opposite vertices
        const rxmesh::VertexHandle p_id = vv[0];
        const rxmesh::VertexHandle r_id = vv[2];
        const rxmesh::VertexHandle q_id = vv[1];
        const rxmesh::VertexHandle s_id = vv[3];

        if (!p_id.is_valid() || !r_id.is_valid() || !q_id.is_valid() ||
            !s_id.is_valid()) {
            return;
        }

        const vec3<T> p(coords(p_id, 0), coords(p_id, 1), coords(p_id, 2));
        const vec3<T> r(coords(r_id, 0), coords(r_id, 1), coords(r_id, 2));
        const vec3<T> q(coords(q_id, 0), coords(q_id, 1), coords(q_id, 2));
        const vec3<T> s(coords(s_id, 0), coords(s_id, 1), coords(s_id, 2));


        T weight = 0;
   weight += dot((p - q), (r - q)) / length(cross(p - q, r - q));
   weight += dot((p - s), (r - s)) / length(cross(p - s, r - s));

        weight /= 2;
        weight = std::max(0.f, weight);

        edge_weights(p_id, r_id) = weight;
        edge_weights(r_id, p_id) = weight;
    };

We now have all the building blocks necessary to perform an iteration step. This is what an iteration loop looks like:

    // process step
    for (int i = 0; i < iterations; i++) {
        // solver for rotation
        calculate_rotation_matrix<float, CUDABlockSize>
            <<<lb_rot.blocks, lb_rot.num_threads, lb_rot.smem_bytes_dyn>>>(
                rx.get_context(),
                ref_vertex_pos,
                deformed_vertex_pos,
                rotations,
                weight_matrix);

        // solve for position
        calculate_b<float, CUDABlockSize>
            <<<lb_b_mat.blocks,
               lb_b_mat.num_threads,
               lb_b_mat.smem_bytes_dyn>>>(rx.get_context(),
                                          ref_vertex_pos,
                                          deformed_vertex_pos,
                                          rotations,
                                          weight_matrix,
                                          b_mat,
                                          constraints);

        laplace_mat.solve(b_mat, *deformed_vertex_pos_mat);
    }

We perform multiple iterations of the algorithm to effectively minimize the energy. We can see in the code above that we alternatively solve for rotation and then for positions. We first calculate the rotation matrix given a set of deformed positions, then calculate the \(b\) matrix (as defined earlier) and then solve the system of equations to identify the new vertex positions.

The following kernel calculates the rotation matrix per vertex to minimize the rigidity energy given their positions.

auto cal_rot = [&](VertexHandle v_id, VertexIterator& vv) {
        Eigen::Matrix3f S = Eigen::Matrix3f::Zero();

        for (int j = 0; j < vv.size(); j++) {

            float w = weight_matrix(v_id, vv[j]);

            Eigen::Vector<float, 3> pi_vector = {
              ref_vertex_pos(v_id, 0) - ref_vertex_pos(vv[j], 0),
              ref_vertex_pos(v_id, 1) - ref_vertex_pos(vv[j], 1),
              ref_vertex_pos(v_id, 2) - ref_vertex_pos(vv[j], 2)
};


            Eigen::Vector<float, 3> pi_dash_vector = {
                deformed_vertex_pos(v_id, 0) - deformed_vertex_pos(vv[j], 0),
                deformed_vertex_pos(v_id, 1) - deformed_vertex_pos(vv[j], 1),
                deformed_vertex_pos(v_id, 2) - deformed_vertex_pos(vv[j], 2)};

            S = S + w * pi_vector * pi_dash_vector.transpose();
        }

        Eigen::Matrix3f U;         // left singular vectors
        Eigen::Matrix3f V;         // right singular vectors
        Eigen::Vector3f sing_val;  // singular values

        svd(S, U, sing_val, V);

        const float smallest_singular_value = sing_val.minCoeff();


        Eigen::Matrix3f R = V * U.transpose();

        if (R.determinant() < 0) {
            U.col(smallest_singular_value) =
                U.col(smallest_singular_value) * -1;
            R = V * U.transpose();
        }

        // Matrix R to vector attribute R

        for (int i = 0; i < 3; i++)
            for (int j = 0; j < 3; j++)
                rotations(v_id, i * 3 + j) = R(i, j);
    };

We first find the Single Value Decomposition (SVD) of the sum (\(S_{i} \)) where

\(S_{i}=\sum_{j\in N(i)} w_{ij}e_{ij}e^{‘T}_{ij} \) (Equation 5 in Sorkine and Alexa(2007))

We then use that to calculate rotation and assign that as the vertice’s rotation. After we calculate rotation, we calculate the right-hand side of Equation 8 as stated earlier.

auto init_lambda = [&](VertexHandle v_id, VertexIterator& vv) {
        // variable to store ith entry of b_mat
        Eigen::Vector3f bi(0.0f, 0.0f, 0.0f);

        // get rotation matrix for ith vertex
        Eigen::Matrix3f Ri = Eigen::Matrix3f::Zero(3, 3);

        for (int i = 0; i < 3; i++) {
            for (int j = 0; j < 3; j++)
                Ri(i, j) = rotations(v_id, i * 3 + j);
        }

        for (int nei_index = 0; nei_index < vv.size(); nei_index++) {
            // get rotation matrix for neightbor j
            Eigen::Matrix3f Rj = Eigen::Matrix3f::Zero(3, 3);
            for (int i = 0; i < 3; i++)
                for (int j = 0; j < 3; j++)
                    Rj(i, j) = rotations(vv[nei_index], i * 3 + j);

            // find rotation addition
            Eigen::Matrix3f rot_add = Ri + Rj;
            // find coord difference
            Eigen::Vector3f vert_diff = {
                ref_vertex_pos(v_id, 0) - ref_vertex_pos(vv[nei_index], 0),
                ref_vertex_pos(v_id, 1) - ref_vertex_pos(vv[nei_index], 1),
                ref_vertex_pos(v_id, 2) - ref_vertex_pos(vv[nei_index], 2)};

            // update bi
            bi = bi +
                 0.5 * weight_mat(v_id, vv[nei_index]) * rot_add * vert_diff;
        }

        if (constraints(v_id, 0) == 0) {
            b_mat(v_id, 0) = bi[0];
            b_mat(v_id, 1) = bi[1];
            b_mat(v_id, 2) = bi[2];
        } else {
            b_mat(v_id, 0) = deformed_vertex_pos(v_id, 0);
            b_mat(v_id, 1) = deformed_vertex_pos(v_id, 1);
            b_mat(v_id, 2) = deformed_vertex_pos(v_id, 2);
        }
    };

After calling both these kernels, we use the solve function to solve the equations. We then update the mesh with this new set of vertex positions.

Results

Using the implementation above, we can create a per-frame ARAP implementation to animate movement for a mesh. In the example below, we fix the upper half vertices of Spot the Cow and allow the bottom half to freely deform. We then move each of the bottom “feet” vertices in every frame to animate a swimming or walking-like movement in Spot.

The bottom feet vertices are user-deformed , the upper vertices are fixed, and the bottom half (except the feet) are free to move—and are determined by the ARAP algorithm.

Here’s another example with the dragon mesh.

Deform a vertex of the dragon’s jaw while keeping its feet fixed.

Further work

Polyscope shows us that the framerate of these demos is very poor (below 30 fps). We would expect it to do better, especially when using the GPU. A performance analysis as well as optimizing of the solver will be required to make the application reach desired levels of real-time processing.

References

[1] Olga Sorkine and Marc Alexa. 2007. As-rigid-as-possible surface modeling. In Proceedings of the fifth Eurographics symposium on Geometry processing (SGP ’07). Eurographics Association, Goslar, DEU, 109–116. https://doi.org/10.2312/SGP/SGP07/109-116

This blog was written by Sachin Kishan during the SGI 2024 Fellowship as one of the outcomes of a two-week project under the mentorship of Ahmed Mahmoud and the support of Supriya Gadi Patil as a teaching assistant.

Author

Leave a Reply

Your email address will not be published. Required fields are marked *