Guénolé Fiche

Research Scientist at @naverlabseurope - Human-centric 3D computer vision

[Guénolé Fiche @GuenoleFiche]

MEGA is the first method to achieve SOTA results in single and multi-output HMR. Want to try it yourself? Code and demo are available at: https://t.co/MDvEx1sRld

Work done in collaboration with @SimonLeglaive , @xavirema , and @fmorenoguer

[Guénolé Fiche @GuenoleFiche]

Reconstructing 3D humans from a single image is highly ambiguous: many 3D poses can explain the same 2D view. Yet, most HMR methods predict a single mesh! 🤯

At #CVPR2025, we present MEGA, a new approach that tackles this challenge. 🧵👇

[Guénolé Fiche @GuenoleFiche]

We propose 2 generation modes:
- In deterministic mode, MEGA predicts all tokens in a single forward pass, ensuring speed and accuracy.
- In stochastic mode, we iteratively sample human mesh tokens, enabling MEGA to produce multiple predictions from a single image.

[Guénolé Fiche @GuenoleFiche]

Work done with @SimonLeglaive , @xavirema , Antonio Agudo, and @fmorenoguer (@eurdigisport , @centralesupelec , @RobotLearnInria , @IRI_robotics )

Paper, code and models: https://t.co/xuMAsgMEUJ (6/6)

[Guénolé Fiche @GuenoleFiche]

How can learned quantized representations help to address human mesh recovery?

In VQ-HPS, to be presented at #ECCV2024, we frame HMR as a classification task in a quantized latent space. (1/6)

[Guénolé Fiche @GuenoleFiche]

VQ-HPS is trained with a cross-entropy loss, without any 3D supervision. Yet, it achieves SOTA results on in-the-wild HMR datasets. Notably, it achieves good performance even in the context of scarce training data. (5/6)