AtomSurf: Surface Representation for Learning on Protein Structures

Vincent Mallet*, Souhaib Attaiki*, Yangyang Miao*, Bruno Correia, Maks Ovsjanikov

ICLR 2025

AtomSurf adapts a state-of-the-art surface-based learning architecture for protein analysis, provides the first direct benchmark comparison against alternative representations, and introduces a novel integrated approach that combines surface and graph-based features to achieve state-of-the-art results.

GANFusion: Feed-Forward Text-to-3D with Diffusion in GAN Space

Souhaib Attaiki, Paul Guerrero, Duygu Ceylan, Niloy Mitra, Maks Ovsjanikov

WACV 2025

GANFusion is a text-guided feed-forward 3D generator that is trained with only single-view image supervision

Unsupervised Representation Learning for Diverse Deformable Shape Collections

Souhaib Attaiki*, Sara Hahner*, Jochen Garcke, Maks Ovsjanikov

3DV 2024

This work introduces an unsupervised mesh autoencoder that creates a universal latent space for diverse 3D shapes using spectral pooling and functional maps, enabling high-quality reconstructions and smooth interpolations without requiring mesh correspondence

Shape Non-rigid Kinematics: A Zero-Shot Method for Non-Rigid Shape Matching via Unsupervised Functional Map Regularized Reconstruction

Souhaib Attaiki and Maks Ovsjanikov

NeurIPS 2023

Zero-shot non-rigid correspondence

Understanding and Improving Features Learned in Deep Functional Maps

Souhaib Attaiki and Maks Ovsjanikov

CVPR 2023 Highlight (top 2.5%)

This work shows that features learned in deep functional maps can be used as pointwise descriptors for direct shape matching and proposes modifications that improve accuracy and bridge intrinsic and extrinsic surface-based learning

Generalizable Local Feature Pre-training for Deformable Shape Analysis

Souhaib Attaiki, Lei Li and Maks Ovsjanikov

CVPR 2023 Highlight (top 2.5%)

VADER investigates the transferability of geometric features for deformable 3D shape analysis and introduces a method for optimizing the receptive field size in transfer learning

SRFeat: Learning Locally Accurate and Globally Consistent Non-Rigid Shape Correspondence

Lei Li, Souhaib Attaiki, Maks Ovsjanikov

3DV 2022

SRFeat introduces a smoothness-regularized contrastive learning framework for non-rigid shape correspondence, combining highly discriminative local features with geometric consistency to improve matching accuracy

NCP: Neural Correspondence Prior for Effective Unsupervised Shape Matching

Souhaib Attaiki and Maks Ovsjanikov

NeurIPS 2022

NCP is a fully unsupervised deep learning-based approach for computing correspondences between 3D shapes, leveraging the inherent structure of neural networks to generate high-quality matches, even in challenging non-isometric or sparse point cloud scenarios

DiffusionNet: Discretization Agnostic Learning on Surfaces

Nicholas Sharp, Souhaib Attaiki, Keenan Crane, Maks Ovsjanikov

SIGGRAPH 2022

Simple and scalable deep learning on meshes, points clouds, etc., via spatial diffusion. The networks automatically generalize across different samplings, resolutions, and even representations. Spatial support is automatically optimized as a parameter!

DPFM: Deep partial functional maps

Souhaib Attaiki, Gautam Pai, Maks Ovsjanikov

3DV 2021 Best Paper Award

DPFM is the first deep learning-based method for computing dense correspondences between non-rigid 3D shapes with strong partiality and can handle both partial-to-full and partial-to-partial matching

Low-light-Image-Enhancement

Python implementation of two low-light image enhancement techniques via illumination map estimation