DPC: Unsupervised Deep Point Correspondence via Cross and Self Construction (3DV 2021)

[Paper] [Introduction Video (2 minutes)] [Introduction Slides] [Full Video (10 minutes)] [Full Slides] [Poster]

This repo is the implementation of DPC. Created by Itai Lang*, Dvir Ginzburg*, Shai Avidan, and Dan Raviv from Tel Aviv University.
*Equal contribution

 

 

Tested environment

• Python 3.6
• PyTorch 1.6
• CUDA 10.2

Lower CUDA and PyTorch versions should work as well.

 

Contents

• Installation
• Datasets
• Train
• Inference
• Cite

 

Installation

Please follow installation.sh or simply run

bash installation.sh 

 

Datasets

The method was evaluated on:

• SURREAL

  • 230k shapes (DPC uses the first 2k).
  • Dataset website
  • This code downloads and preprocesses SURREAL automatically.

• SHREC’19

  • 44 Human scans.
  • Dataset website
  • This code downloads and preprocesses SURREAL automatically.

• SMAL

  • 10000 animal models (2000 models per animal, 5 animals).
  • Dataset website
  • Due to licencing concerns, you should register to SMAL and download the dataset.
  • You should follow data/generate_smal.md after downloading the dataset.
  • To ease the usage of this benchmark, the processed dataset can be downloaded from here. Please extract and put under data/datasets/smal

• TOSCA

  • 41 Animal figures.
  • Dataset website
  • This code downloads and preprocesses TOSCA automatically.
  • To ease the usage of this benchmark, the processed dataset can be downloaded from here. Please extract and put under data/datasets/tosca

 

Training

For training run

python train_point_corr.py --dataset_name <surreal/tosca/shrec/smal>

The code is based on PyTorch-Lightning, all PL hyperparameters are supported. (limit_train/val/test_batches, check_val_every_n_epoch etc.)

 

Tensorboard support

All metrics are being logged automatically and stored in

output/shape_corr/DeepPointCorr/arch_DeepPointCorr/dataset_name_<name>/run_<num>

Run tesnroboard --logdir=<path> to see the the logs.

Example of tensorboard output:

 

Inference

For testing, simply add --do_train false flag, followed by --resume_from_checkpoint with the relevant checkpoint.

python train_point_corr.py --do_train false  --resume_from_checkpoint <path>

Test phase visualizes each sample, for faster inference pass --show_vis false.

We provide a trained checkpoint repreducing the results provided in the paper, to test and visualize the model run

python train_point_corr.py --show_vis --do_train false --resume_from_checkpoint data/ckpts/surreal_ckpt.ckpt

 

Citing & Authors

If you find this repository helpful feel free to cite our publication -

@InProceedings{lang2021dpc,
  author = {Lang, Itai and Ginzburg, Dvir and Avidan, Shai and Raviv, Dan},
  title = {{DPC: Unsupervised Deep Point Correspondence via Cross and Self Construction}},
  booktitle = {Proceedings of the International Conference on 3D Vision (3DV)},
  pages = {1442--1451},
  year = {2021}
}

Contact: Dvir Ginzburg, Itai Lang.