Generation3D

3D Shape Generation Baselines in PyTorch.

Feature

• Hack of DataParallel for balanced memory usage
• More Models WIP
• Configurable model parameters
• Customizable model, dataset

Representation

• 💎 Polygonal Mesh
• 👾 Volumetric
• 🎲 Point Cloud
• 🎯 Implicit Function
• 💊 Primitive

Input Observation

• 🏞 RGB Image
• 📡 Depth Image
• 👾 Voxel
• 🎲 Point Cloud
• 🎰 Unconditional Random

Evaluation Metrics

• Chamfer Distance
• F-score
• IoU

Model Zoo

• 💎 Pixel2Mesh
• 🎯 DISN
• 👾 3DGAN
• 👾 Voxel Based Method
• 🎲 PointCloud Based Method

Get Started

Environment

• Ubuntu 16.04 / 18.04
• Pytorch 1.3.1
• CUDA 10
• conda > 4.6.2

Using Anaconda to install all dependences.

conda env create -f environment.yml

Train

CUDA_VISIBLE_DEVICES=<gpus> python train.py --options <config>

Predict

CUDA_VISIBLE_DEVICES=<gpus> python predictor.py --options <config>

Evaluation [WIP]

Custom guide

• custom scheduler for training/inference loop, add code in scheduler and inherit base class.
• custom model in models/zoo
• custom config options in utils/config
• custom dataset in datasets/data

External

• Chamfer Distance

Baselines

Pixel2Mesh 🏞 💎

• Input: RGB Image
• Representation: Mesh
• Output: Mesh _camera-view

DISN 🏞 🎯

• Input: RGB Image
• Representation: SDF
• Post-processing: Marching Cube
• Output: Mesh _camera-view

3DGAN 🎰 👾

• Input: Random Noise
• Representation: Volumetric
• Output: Voxel

Acknowledgements

Our work is based on the codebase of an unofficial pixel2mesh framework. The Chamfer loss code is based on ChamferDistancePytorch.

Official baseline code

• DISN: Deep Implicit Surface Network for High-quality Single-view 3D Reconstruction

• Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images

• Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling

License

Please follow the License of official implementation for each model.