From a22e667a41606770195caeab1fde91055f941fdb Mon Sep 17 00:00:00 2001 From: schwabjohannes Date: Thu, 19 Dec 2024 08:02:20 +0100 Subject: [PATCH] Update README.md --- README.md | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/README.md b/README.md index f5a4853..350e499 100644 --- a/README.md +++ b/README.md @@ -34,6 +34,14 @@ There are many more parameters that define the neural network that in normal use It is also possible to follow the training process by looking at the tensorboard that is updated after every epoch. +### Training the VAE with a rigid body prior + +A modified version of Dynamight estimates rigid deformations for user defined bodies. In this version, the Encoder is exactly the same, whereas the decoder predicts a rotation and translation for each body that is specified by binary masks (--deformation-masks). For a given latent representation $z$ and $N$ masks $\{m_1,\ldots, m_N\}$ is of the form + +$$ D_z(x) = \sum_{i=1}^N \sum_{j=1}^{N_i} \varphi_j(R_{i,z}(x)),$$ + +where $N_i$ is the number of Gaussians within mask $m_i$ and $R_{i,z}ยง is the rigid transform that is estimated for mask $m_i$ from the latent representation $z$. + ### Visualization