LONG-XI / Geo-CNN

Modeling Local Geometric Structure of 3D Point Clouds using Geo-CNN

Bibtex

@article{DBLP:journals/corr/abs-1811-07782,
  author    = {Shiyi Lan and
              Ruichi Yu and
              Gang Yu and
              Larry S. Davis},
  title     = {Modeling Local Geometric Structure of 3D Point Clouds using Geo-CNN},
  journal   = {CoRR},
  volume    = {abs/1811.07782},
  year      = {2018},
  url       = {http://arxiv.org/abs/1811.07782},
  archivePrefix = {arXiv},
  eprint    = {1811.07782},
  timestamp = {Mon, 26 Nov 2018 12:52:45 +0100},
  biburl    = {https://dblp.org/rec/bib/journals/corr/abs-1811-07782},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Installation and Usage

We re-implemented the Geo-CNN following Frustum PointNets.

Follow the instruction of installing Frustum PointNets and thus compile Geo-Conv operator located at models/tf_ops/geoconv.

Use scripts/command_train_geocnn_v1.sh and command_test_geocnn_v1.sh to train/test Geo-CNN.

TODO

• Combine GeoCNN and PointNet++
• GeoCNN on other 3D datasets (ModelNet40, ScanNet)

Others

• Well-trained parameters
• This implementation is slightly different from the original version on a private deep learning architure.