We propose two gradient-based deep methods to perform an image clustering invariant to spatial, color and morphological transformations. We showcase its robustness and interpretability by visualizing clustering results over real photograph collections and instagram hashtags.
@inproceedings{monnier2020dticlustering,
title={Deep Transformation-Invariant Clustering},
author={Monnier, Tom and Groueix, Thibault and Aubry, Mathieu},
booktitle={arXiv:2006.11132},
year={2020}
}
A combination of learned elementary 3D structures are used to represent 3D shapes. The learned elementary 3D structures are highly interpretable and lead to clear improvements in 3D shape generation and matching.
@inproceedings{deprelle2019learning,
title={Learning elementary structures for 3D shape generation and matching},
author={Deprelle, Theo and Groueix, Thibault and Fisher, Matthew and Kim, Vladimir G and Russell, Bryan C and Aubry, Mathieu},
booktitle={Neurips},
year={2019}
}
We propose a solution to the task of putting objects of arbitrary topologies in correspondences. Matching can be solved by predicting dense deformations. In this work, we introduce a deep learning method to deform any shape into any other, and a cycle-consistent loss on pointclouds to regularize the learned deformations.
@inproceedings{groueix19cycleconsistentdeformation,
title = {Unsupervised cycle-consistent deformation for shape matching},
author = {Groueix, Thibault and Fisher, Matthew and Kim, Vladimir G. and Russell, Bryan and Aubry, Mathieu},
booktitle = {Symposium on Geometry Processing (SGP)},
year = {2019}
}
Given two input shapes without vertex correspondences, the goal is to establish correspondences between them. To do so, we learn a smooth deformation transforming a template shape into the input shape. The two reconstructions are naturally in correspondences, as they span from two different deformations of the same template.
@inproceedings{groueix2018b,
title = {3D-CODED : 3D Correspondences by Deep Deformation},
author={Groueix, Thibault and Fisher, Matthew and Kim, Vladimir G. and Russell, Bryan and Aubry, Mathieu},
booktitle = {ECCV},
year = {2018}
}
3D Synthesis : Instead of generating point clouds or voxels, AtlasNet directly generates meshes at arbitrary resolution. It has strong advantages : generalization capabilities, morphing, co-segmentation, regular texture parameterization. We demonstrate its strength on an autoencoder, and single-view reconstruction from one still image.
@inproceedings{groueix2018,
title={{AtlasNet: A Papier-M\^ach\'e Approach to Learning 3D Surface Generation}},
author={Groueix, Thibault and Fisher, Matthew and Kim, Vladimir G. and Russell, Bryan and Aubry, Mathieu},
booktitle={Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR)},
year={2018}
}
Approximate real-time previzualition of Global Illumination (GI) relying on NVIDIA Optix (parallelized ray tracing). To approximate GI in real time, we apply a Joint Bilateral Filter on a sub-sampled rendered image and a high-resolution G-buffer (containing depth, normal, and albedo).
@inproceedings {egp.20161048,
booktitle = {EG 2016 - Posters},
editor = {Luis Gonzaga Magalhaes and Rafal Mantiuk},
title = {{Interactive Monte-Carlo Ray-Tracing Upsampling}},
author = {Boughida, Malik and Groueix, Thibault and Boubekeur, Tamy},
year = {2016},
publisher = {The Eurographics Association},
ISSN = {1017-4656},
DOI = {10.2312/egp.20161048}
}
@inproceedings{groueix2018,
title={{Learning 3D Generation and Matching}},
author={Groueix, Thibault},
school={Ecole Nationale des Ponts et Chaussees},
year={2020}
}
