Surface class segmentation in CAD models with MeshCNN

This project modifies MeshCNN to handle large meshes efficiently and applies the improved framework to learn surface type segmentation on ABC Dataset, a large dataset of CAD models.

Project by Andrés Mandado

Getting Started

Installation

• Clone this repo:

git clone https://github.com/atomicsulfate/meshcnn-4-cadseg.git
cd meshcnn-4-cadseg.git

• Install dependencies with conda (creates an environment called cgp-meshcnn-basic):

cd development
conda env create -f basic_environment.yml

ABC Dataset surface class segmentation

Download the dataset

bash ./scripts/seg/get_10K_dataset.sh

Run training (if using conda env first activate env e.g. conda activate cgp-meshcnn-basic)

bash ./scripts/seg/train.sh

To view the training loss plots, in another terminal run tensorboard --logdir runs and click http://localhost:6006.

Run test and export the intermediate pooled meshes:

bash ./scripts/seg/test.sh

Visualize the network-learned segmentation vs labels:

bash ./scripts/seg/view.sh

Some segmentation result examples:

Note, you can also get pre-trained weights using bash ./scripts/seg/get_pretrained.sh.

Project structure

• development
  • meshcnn: Original meshcnn code (unchanged).
  • data, models: meshcnn extensions (e.g. sparse pooling, distributed training).
  • test.py, train.py: Test and train scripts.
  • scripts
    • prepro: Scripts for data preprocessing, synthetic sample generation, mesh visualization...
    • seg: Scripts to execute basic workflow: train,test,view results.
• docs
  • website