Survey and Evaluation of Neural 3D Shape Classification Approaches

Abstract

<div> <div> <div> <p>Classification of 3D objects – the selection of a category in which each object belongs – is of great interest in the field of machine learning. Numerous researchers use deep neural networks to address this problem, altering the network architecture and representation of the 3D shape used as an input. To investigate the effectiveness of their approaches, we conduct an extensive survey of existing methods and identify common ideas by which we categorize them into a taxonomy. Second, we evaluate 11 selected classification networks on three 3D object datasets, extending the evaluation to a larger dataset on which most of the selected approaches have not been tested yet. For this, we provide a framework for converting shapes from common 3D mesh formats into formats native to each network, and for training and evaluating different classification approaches on this data. Despite being generally unable to reach the accuracies reported in the original papers, we can compare the relative performance of the approaches as well as their performance when changing datasets as the only variable to provide valuable insights into performance on different kinds of data. We make our code available to simplify running training experiments with multiple neural networks with different prerequisites. </p> </div> </div> </div>

Deep Synthesis of Cloud Lighting

“…Satilmis et al [5] and Mirbauer et al [6] proposed methods to generate whole-sky cloudy lighting which is suitable for use as an environment map and therefore can be directly integrated into rendering systems. Satilmis et al [5] used U-net structured autoencoders to generate cloudy sky lighting from encoded clear sky lighting and a cloud mask.…”

“…Satilmis et al [5] used U-net structured autoencoders to generate cloudy sky lighting from encoded clear sky lighting and a cloud mask. Mirbauer et al [6] used generative adversarial networks to synthesize cloudy skies conditioned on sun position and cloud coverage.…”

“…Static cloud synthesis on the hemisphere for use in rendering can be modeled by existing techniques such as Satilmis et al [5] and Mirbauer et al [6]. However in order to generate realistic cloud movement over multiple frames, we can start with these approaches or captured environment illumination and generate a sequence of images which contain cloud movement.…”

“…To overcome these limitations, recently generative machine learning methods have been developed [5], [6] to generate cloudy sky environment maps. These typically are trained on images of real cloudy skies and are combined with an analytical sky model input to generate clouds either via a Generative Adversarial Network (GAN) as in the work by Mirbauer et al [6], or via user or artist specified masks and a U-net as used in Satilmis et al [5]. These methods can generate realistic images of cloudy skies suitable for use as environment maps without the complexities of representing and rendering cloud volumes.…”

“…Our approach is trained and validated using a database of captured sequences of cloud movement, and we show how this can be applied to either a single captured cloudy environment image, or an image generated by a generative cloud model [5], [6]. Our technique allows animation of artistically generated clouds and produces results which can be directly used in a rendering system for environment illumination.…”