Point Cloud Autoencoder

A Jupyter notebook containing a PyTorch implementation of Point Cloud Autoencoder inspired from "Learning Representations and Generative Models For 3D Point Clouds". Encoder is a PointNet model with 3 1-D convolutional layers, each followed by a ReLU and batch-normalization. Decoder is a MLP with 3 fully connected layers with ReLU activations except the output layer.

Data

Dataset is sampled from "A Scalable Active Framework for Region Annotation in 3D Shape Collections". 3746 point cloud models in "chair" category are randomly downsampled to 1024 points. The result is saved as a numpy array with 3746x1024x3 dimensions. You can use your own data by feeding the pc_array with a numpy array using the same data format.

Results

Epoch 1
Epoch 500
Epoch 1000

Environment

Tested with:
PyTorch v1.5.0  
PyTorch3D v0.2  
Nvidia RTX2070 with CUDA 10.2