Sharma et al. use a volumetric convolutional denoising auto-encoder for shape completion and classification on the ModelNet  Dataset. Their approach is comparably simple — the extension of the regular denoising auto-encoder to volumetric data is straight forwards on the used resolution of $30^3$. The architecture is quite simple and illustrated in Figure 1. A dropout layer after the input simulates noise and a bottleneck layer of dimensionality $6912$ is used.
Figure 1 (click to enlarge): The volumetric, convolutional denoising auto-encoder architecture used by Sharma et al.
In experiments on the ModelNet  Dataset, they demonstrate the applicability of their model for classification and shape completion. For classification they outperform the ShapeNet model  both when training an SVM on the $6912$ dimensional representation and when fine-tuning by adding two additional fully connected layers. However, VoxNet  still outperforms their approach. Qualitative results for shape completion on random noise and slicing noise are shown in Figure 2 and 3, respectively. It seems as if the model struggles most with the low resolution. Especially for slicing noise, the model performs poorly because of the low resolution.
Figure 2 (click to enlarge): Qualitative results for shape completion from random noise. Comparison to ShapeNet .
Figure 3 (click to enlarge): Qualitative results of shape completion on slicing noise and comparison to ShapeNet .
What is your opinion on the summarized work? Or do you know related work that is of interest? Let me know your thoughts in the comments below or using the following platforms: