A Unified Point-Based Framework for 3D Segmentation

This repository contains the implementation of A Unified Point-Based Framework for 3D indoor scene semantic segmentation. Idea

Prepare Data and 3D Evaluation

The preprocessing steps of ScanNet v2 dataset. Some of the codes are borrowed from 3DMV. We render the images from 3D indoor scenes in ScanNetV2 dataset, and use the images to train deeplab.

DeepLab

The implementation is orignal from DeepLab. We use DeepLab as our image feature extractor, and back-project the feature onto 3D vertices. To train or to finetune our model, please see the README.md inside the directory.

Joint Model

The model predicts points' semantics with a local encoder, a global encoder and a decoder. The input features (color feature, geometry feature) are fused into point-based representation. To train or to finetune our model, please see the README.md inside the directory.

3D Evaluation

For evaluation, we use the code provided from ScanNet Banchmark to evaluate our model performance.

Citation

If you find this repository useful, please cite:

@article{chiang2019unified,
  title={A Unified Point-Based Framework for 3D Segmentation},
  author={Chiang, Hungyueh and Lin, Yenliang and Liu, Yuehcheng and Hsu, Winston H},
  journal={arXiv preprint arXiv:1908.00478},
  year={2019}
}