A Differentiability of $\mathbf{T}_g$ with respect to $r$

In this appendix we study continuity and differentiability of \(\mathbf{T}_g\) (i.e. of \(\mathbf{R}_g,\mathbf{t}_g\)) with respect to \(r\). Variations in \(r\) imply a smooth variation on the domains \(\Omega_g\) being considered for the Procrustes analysis (domains are enlarged or reduced by boundary \(\partial \Omega_g\)). Regarding the continuous Procrustes problem definition in (Al-Aifari et al., 2013), we can define our analogous space-continuous squared Procrustes residual for \(\mathbf{X}_g\) and \(\mathbf{Y}_g\) as: \[ d^2_{\mathcal{P}}(\mathbf{X}_g,\mathbf{Y}_g)=\begin{matrix} \\ \min \\ \mathbf{R}_g,\mathbf{t}_g \end{matrix}\left(\int_{\Omega_g}\left \|\mathbf{R}_g\mathbf{x}+\mathbf{t}_g-\mathbf{y}\right\|^2 \mathrm{d} {\Omega}\right), \tag{25} \] where \(\mathbf{x}\) and \(\mathbf{y}\) represent mapped points of the shapes' surface manifolds. For now on, as \(\mathbf{x}\) are mapped to \(\mathbf{y}\) with area preserving maps (Melzi et al., 2019), we will use \(\mathrm{d}\Omega_g\) to refer to the differential elements on both the current and the target shape manifolds. The optimal translation, in this continuous formulation, is \[ \mathbf{t}_g=\frac{\int_{\Omega_g}\mathbf{y}\mathrm{d}\Omega_g}{\int_{\Omega_g}\mathrm{d}\Omega_g}-\frac{\int_{\Omega_g}\mathbf{R}_g\mathbf{x}\mathrm{d}\Omega_g}{\int_{\Omega_g}\mathrm{d}\Omega_g}=\frac{\int_{\Omega_g} (\mathbf{y}-\mathbf{R}_g\mathbf{x})\mathrm{d}\Omega_g}{\int_{\Omega_g} \mathrm{d}\Omega_g}, \] and can be differentiated with respect to \(r\) to obtain: \[ \frac{\partial}{\partial r}\mathbf{t}_g=\frac{\partial}{\partial r}\left(\frac{\int_{\Omega_g} (\mathbf{y}-\mathbf{R}_g\mathbf{x})\mathrm{d}\Omega_g}{\int_{\Omega_g} \mathrm{d}\Omega_g} \right ) \\ =\frac{\partial}{\partial r}\left(\frac{1}{A} \right )\int_{\Omega_g} (\mathbf{y}-\mathbf{R}_g\mathbf{x})\mathrm{d}\Omega_g+\frac{1}{A}\frac{\partial}{\partial r}\left(\int_{\Omega_g} (\mathbf{y}-\mathbf{R}_g\mathbf{x})\mathrm{d}\Omega_g\right) \\ =-\frac{1}{A^2}\frac{\partial A}{\partial r}\int_{\Omega_g} (\mathbf{y}-\mathbf{R}_g\mathbf{x})\mathrm{d}\Omega_g \\ +\frac{1}{A}\left( \int_{\Omega_g}\frac{\partial \mathbf{R}_g}{\partial r}\mathbf{x}\mathrm{d}\Omega_g +\int_{\partial\Omega_g} (\mathbf{y}-\mathbf{R}_g\mathbf{x})\mathrm{d}(\partial\Omega_g)\right), \tag{26} \] where \(A(r)\) is the area of the open domain \(\Omega_g\) and, in the last step, we applied the Leibniz's integral rule. Note that (26) depends on the existence of \(\frac{\partial \mathbf{R}_g}{\partial r}\). The optimal Procrustes rotation component \(\mathbf{R}_g\) can be obtained from: \[ \mathbf{R}_g=\sqrt{\mathbf{M}_g\mathbf{M}_g^\intercal}\,\mathbf{M}_g^{-1}, \tag{27} \] where \(\mathbf{M}_g(r)\in\mathbb{R}^{3\times3}\) is equal to: \[ \mathbf{M}_g(r)=\int_{\Omega_g}\left(\mathbf{x}-\bar{\mathbf{x}}_g\right)\left(\mathbf{y}-\bar{\mathbf{y}}_g\right)^\intercal\mathrm{d}\Omega_g. \tag{28} \] In this continuous formulation \(\bar{\mathbf{x}}_g={\int_{\Omega_g}\mathbf{x}\mathrm{d}\Omega_g}/{\int_{\Omega_g}\mathrm{d}\Omega_g}\) and \(\bar{\mathbf{y}}_g={\int_{\Omega_g}\mathbf{y}\mathrm{d}\Omega_g}/{\int_{\Omega_g}\mathrm{d}\Omega_g}\). Equation (27), and therefore \(\mathbf{R}_g\), is continuous differentiable with respect to \(r\) given two conditions: \(\mathbf{M}_g\) being continuous differentiable and existence of \(\mathbf{M}^{-1}_g\) for \(r\in[r_0,R(t)]\). \(\mathbf{M}_g(r)\) is differentiable with respect to \(r, \forall \, r>0\): \[ \frac{\partial}{\partial r}\mathbf{M}_g= \int_{\Omega_g}\frac{\partial}{\partial r}\left(\left(\mathbf{x}-\bar{\mathbf{x}}_g\right)\left(\mathbf{y}-\bar{\mathbf{y}}_g\right)^\intercal\right)\mathrm{d}\Omega_g \\ + \int_{\partial\Omega_g}\left(\mathbf{x}-\bar{\mathbf{x}}_g\right)\left(\mathbf{y}-\bar{\mathbf{y}}_g\right)^\intercal\mathrm{d}(\partial\Omega_g). \tag{29} \] On the other hand, \(\mathbf{M}_g^{-1}\) exists if there exist at least three non-aligned points \(\mathbf{x}\) and three non-aligned points \(\mathbf{y}\) in the domains \(\Omega_g\) of each manifold. This condition is certainly achieved for any 3D object (planar or volumetric) and \(r>0\).
Ignacio Cuiral-Zueco
Ignacio Cuiral-Zueco received the Industrial Engineering Master degree majoring in robotics and computer vision in the University of Zaragoza, Spain, in 2019. He is currently enrolled in the Ph.D. program in systems engineering and computer science from the University of Zaragoza. He is member of the Robotics, Perception and Real-Time research group. His current research interests include computer vision, control engineering and robotics.
Gonzalo López-Nicolás
Gonzalo López-Nicolás (SM'15) received the Ph.D. degree in Systems Engineering and Computer Science from Universidad de Zaragoza, Spain, in 2008. He is currently Full Professor of the Departamento de Informática e Ingeniería de Sistemas, Universidad de Zaragoza. He is a member of the Instituto de Investigación en Ingeniería de Aragón (I3A). His current research interests are focused on shape control, visual control, multi-robot systems, and application of computer vision to robotics.