Correspondence Learning via Linearly-invariant Embedding

This repository is the official implementation of Correspondence Learning via Linearly-invariant Embedding

This code was written by Marie-Julie Rakotosaona and Riccardo Marin.

NEWS: A PyTorch implementation of the method can be found here

Requirements

Prerequisites:

• CUDA and CuDNN
• Python 2.7
• Tensorflow 1.15

Pre-trained Models

You can download pretrained models here:

• Pretrained model

Download pretrained models from code:

cd pretrained_models
python download_models.py

Dataset

• Our dataset

Download dataset from code:

cd data
python download_data.py

Training

To train the basis model:

python src/train_basis.py --log_dir "path to basis model directory"

To train the transformation model:

python src/train_descriptors.py --log_dir "path to transformation model directory" --model_path "path to basis model"

Evaluation

To evaluate a model, you can use the "test.py" file. For example, to evaluate the pre-trained model on the Noise dataset:

Run pretrained model on FAUST testset:

python ./src/test.py --i FAUST_noise_0.01 -n log_model --epoch=4427

Results

The evaluation of the correspondence for point clouds generated from the FAUST dataset without or with additional noise as mean error, with cumulative curves and a qualitative example.

Method	No noise	Noise
our	5.4e-2	6.6e-2
GFM	2.9e-1	3.4e-1
Uni20	7.5e-2	8.5e-2
Uni60	6.9e-2	8.1e-2
3DC	7.0e-2	7.3e-2
FMAP	1.3e-1	1.4e-1
FMAP+ZOO	1.1e-1	1.3e-1
GFM+ZOO	3.1e-1	3.8e-1

To replicate our evaluation, you can run the "evaluation.m" in Matlab. It is already configurated to run on the output of pre-trained model. In the paper we computed the geodesic distance matrices using an old pre-compiled .mex file for fastmarch algorithm. For the moment (and for the sake of semplicity), in this git we attached a pre-computed geodesic distance matrix used for all the pairs. It is an approximation, but this will let you to run the evaluation. This should return to you a plot like this:

Citation

If you use our work, please cite our paper.

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. For any commercial uses or derivatives, please contact us.