Today neural networks successfully solve the problem of 2D and 3D image segmentation. Such approaches take a 2D or 3D array as input, and produce a binary pixel or voxel image as an output. The segmentation problem can also be solved on surface meshes of 3D objects. In this case, the segmentation method takes as input the vertices and edges of the surface mesh of an object, and at the output returns an integer value of the class for each vertex or edge of the provided mesh [1]. Surface mesh segmentation tasks are relevant in biomedical applications and in computer graphics. Both these fields suffer from the same problem, which is the lack of relevant data. In our work, we propose a new approach for supervised surface mesh segmentation that can efficiently work with small datasets. We represent 3D geometry using signed distance function, and encode local geometric features to use them as an input for the convolutional neural network. The performance of our method is demonstrated on surface meshes of human heart ventricles from the ACDC Challenge dataset [2]. The most important advantage of our approach is its ability to learn on few examples. It is also able to handle thin and closely spaced surfaces, as it works with the geometry features and does not depend on the size of the surface mesh elements. Furthermore, the required accuracy of our method can be achieved with 15 times less trainable parameters than it is necessary for PointNet on the same dataset.