TreeGCN-ED

An autoencoder for point cloud encoding-decoding build using tree-GAN as base work.

[Paper Link]

Dataset Generation Step

• ShapeNetBenchmarkV2 dataset is used.
• To sample pointcloud from mesh:
  • https://medium.com/@daviddelaiglesiacastro/3d-point-cloud-generation-from-3d-triangular-mesh-bbb602ecf238
  • Area of triangle with 3 coordinates: https://math.stackexchange.com/a/1951650
  • https://www.youtube.com/watch?v=HYAgJN3x4GA
• Sample point cloud is spherical normalized.
• Data generation code is present in Preprocessing_Data folder.
• ShapeNetBenchmarkV2 numpy format dataset: Link

Pre-trained model

• Download pre-trained model from google drive:
  • treeED_ckpt
  • treeED_eckpt
• Keep treeED_ckpt, treeED_eckpt as it is in code directory.

Results

Intra-class interpolation results
Chair to Chair	Table to Table	Airplane to Airplane
Inter-class interpolation results
Laptop to Airplane	Cup to Table	Car to Chair

Reference

[1] 3D Point Cloud Generative Adversarial Network Based on Tree Structured Graph Convolutions [ Dong Wook Shu, Sung Woo Park, Junseok Kwon ]

Bibtex Citation

@inproceedings {10.2312:pg.20231278,
booktitle = {Pacific Graphics Short Papers and Posters},
editor = {Chaine, Raphaëlle and Deng, Zhigang and Kim, Min H.},
title = {{TreeGCN-ED: A Tree-Structured Graph-Based Autoencoder Framework For Point Cloud Processing}},
author = {Singh, Prajwal and Tiwari, Ashish and Sadekar, Kaustubh and Raman, Shanmuganathan},
year = {2023},
publisher = {The Eurographics Association},
ISBN = {978-3-03868-234-9},
DOI = {10.2312/pg.20231278}
}