Corentin Dumery Evocube Pattern Making B-Mesh Modeller Aesthetic functions Cow texture generator Resume Gallery Blog

Corentin Dumery

PhD student at  EPFL
CVLab

Contact:
first name dot last name at epfl dot ch

My research interests lie at the intersection of computer vision and computer graphics. I am dedicated to advancing machine perception through 3D scene reconstruction and understanding, enabling machines to not only see their environment but also comprehend and interact with it. My work also emphasizes 3D content creation for digital AR/VR environments, leveraging both real-world reconstruction and AI-assisted 3D generation.

I graduated from NUS and Télécom Paris in 2020. After graduation, I worked at CEA Paris-Saclay on polycube mapping, and later joined Prof. Olga Sorkine-Hornung's lab at ETH Zürich as a visiting researcher on garment fabrication.

For those who are unsure, here's how my name is pronounced in French.

⟶ Students interested in a semester project or master thesis, please consult our lab's project page.
⟶ Master/Bachelor summer interns, please apply to the Summer@EPFL program.
⟶ PhD applicants, please refer to the doctoral program page or our student-wiki.

If you would like to contact me, please make it clear you have had a look at these resourses and avoid sending a generic email.

3D Garment Reconstruction [CVPR24, NeurIPS24]

From a single 2D image or from a partial point cloud, how can we recover a full 3D garment?
Recovering a full 3D model of a garment from just a single image, especially when it's being worn or handled, can be a complex task due to partial visibility. To achieve this, we leverage garment-specific geometric priors to accurately reconstruct the visible parts of garments while inferring the unseen areas to recover a complete and realistic 3D model.

Evocube [CGF22]

Given an input 3D shape, how can we deform it into a similar polycube?
→ check out the Project page. Polycubes, i.e. aggregates of axis-aligned cubes, are most notably used for all-hex meshing, but their constrained nature and topology conditions make polycube parameterization a challenging task. Previous work has shown the feasibility of precomputing polycube orientation directly on the input mesh, effectively labeling each boundary face in blue/white/red representing the XYZ axes.
How can we derive a fitting polycube topology from a triangle mesh? We propose the use of a novel genetic algorithm in the context of polycube labeling optimization, defining fitness, crossover, and mutations.

Computational Pattern Making [SIGGRAPH22]

Given a 3D garment, how can we generate a 2D sewing pattern that can effectively be used to manufacture it? Our algorithm segments an input 3D shape into patches and computes their 2D parameterization, while accounting for the unique properties and constraints of tailoring, such as seam symmetry, darts, fabric grain alignment, and a flattening distortion measure that models woven fabric deformation, respecting its anisotropic behavior.

B-Mesh Modeller

With two friends from Télécom Paris, we created a 3D modelling software based on a novel approach described in a research paper. The idea is to create an initial mesh in only a few minutes by placing spheres in 3D to represent the skeleton of the modelled object. The user can freely create and customize the mesh. It can be modified in real-time. (link to the project)




Aesthetic functions

This is a fun little project to explore the artistic side of two-dimensional functions. All you have to do is enter a math function and play with the sliders to generate some stunning artworks. Any function works, no matter how complex, but even on simpler ones the results are often surprising. If you do try it out, please send me your creations! (link to the project)

Black hole of odd dimension Sea floor
Color dance Miracle of Life


And some of them move, too! There's plenty more where these come from, some moving here, some surprising there, and they all have something that all the others don't have. I always love receiving creations from other people, so feel free to send me an email if you find inspiration!

Cow Texture generator

Ever felt the need to have an infinite supply of cow pattern textures?

This project was inspired by The leopard never changes its spots, a SIGGRAPH 2020 paper by Malheiros, Marcelo de G. and Fensterseifer, Henrique and Walter, Marcelo. This paper uses a reaction-diffusion model to approximate tissue growth, and successfully generates a few 2D patterns matching real species. This project aims to adapt this model to generate cow patterns, which were not covered in the original article. (link to the project)


Evaluation of a Spectral Data Transformation Method for Meaningful Mesh Segmentation

I wanted to see if I could transform a 3D mesh in a weaningful way to make 3D segmentation easier. To evaluate this transformation, I used this awesome dataset which includes human-generated ground truth segmentations, and used simple clustering algorithms to generate a segmentation. By comparing results with the ground truth, I was able to measure the efficiency of this approach and identify the circumstances under which it's useful. For more detail, here is a link to the complete study. You can also take a look at the project on github.

Design of Implants for Skull Reconstructive Surgery

This project aims to make 3D skull implant generation as easy as possible. The idea is to input a mesh derived from a CT scan and use two edge loops to specify the part of the skull where an implant should be generated. The skull layer is then reconstructed to make a perfect-fit implant. Then, a flattening algorithm is used to flatten each layer of the implant to make it suitable for implant 3D printing. Cutting path are added to release some flattening constraints. This work was done for the National University of Singapore in collaboration with Osteopore, and uses a patented software from NUS to reconstruct the outer layer of the skull using symmetry constraints.

Potato Generator

Simple project that generates a 3D potatoïd based on input parameters. Perturbations with a given frequency, amplitude and direction are applied to a sphere to create the illusion of a natural object. (link to the project)

[FRENCH] Modélisation agricole et optimisation de la répartition des surfaces

Here's a video I made when I was 19 showing the different steps in my research trying to optimize the area allocation of an agricultural exploitation. It's in French and quite simple since this was done very early during my studies, but if you have any questions I'd be more than willing to speak about it with you. It was a fun experience and I enjoyed the freedom that I was given on this project.

More information

For more information, feel free to contact me at corentin.dumery at gmail.com. You can also download my resume here.