Roman Klokov

Roman Klokov

Postdoctoral researcher working on deep learning for geometric data.

Publications

You can also find a complete list of my publications on my Google Scholar profile.

Conference Papers

Self-Supervised Dual Contouring

Published in Conference on Computer Vision and Pattern Recognition, 2024

The paper proposes a novel training strategy for neural dual contouring differentiable meshing model based on establishing consistency between input ground truth signed distance function (SDF) values/normals and SDF values/normals to the predicted mesh. The method can additionally be used to regularize predictions of neural SDF models.

R. Sundararaman, R. Klokov and M. Ovsjanikov. "Self-Supervised Dual Contouring." In CVPR'24.

Download Paper

VoroMesh: Learning Watertight Surface Meshes with Voronoi Diagrams

Published in International Conference on Computer Vision, 2023

The paper proposes a novel differentiable resperentation for surfaces based on Voronoi diagrams with direct access to watertight mesh extraction. VoroMesh is verified in two settings: direct per-object optimization (overfitting), and infernce-based mesh reconstuction from low-resolution grids of signed distance function values.

N. Maruani, R. Klokov, M. Ovsjanikov, P. Alliez and M. Desbrun. "VoroMesh: Learning Watertight Surface Meshes with Voronoi Diagrams." In ICCV'23.

Download Paper | Paper Code

Discrete Point Flow Networks for Efficient Point Cloud Generation

Published in European Conference on Computer Vision, 2020

The paper proposes a novel hierarchical probabilistic generative model for point clouds based on invertible conditional affine coupling flows reimagined for point clouds, and capable of generation of variable size point clouds. The model was adapted to single-view reconstruction task and demonstrated performance similar to the state of the art while being two magnitudes faster during training and sampling.

R. Klokov, E. Boyer and J. Verbeek. "Discrete Point Flow Networks for Efficient Point Cloud Generation." In ECCV'20.

Download Paper | Paper Code

Probabilistic Reconstruction Networks for 3D Shape Inference from a Single Image

Published in British Machine Vision Conference, 2019

The paper analyses learning strategies for the single-view reconstruction task. The problem is framed as a conditional generation task with image-conditioned prior. Various learning strategies and image conditioning mechanisms are compared to show that variational training scheme with shape-conditioned posterior and image-conditined prior distributions for the global latent variable.
This work recieved a Best Science Paper Honorable Mention Award.

R. Klokov, J. Verbeek and E. Boyer. "Probabilistic Reconstruction Networks for 3D Shape Inference from a Single Image." In BMVC'19.

Download Paper | Paper Code

Escape from Cells: Deep Kd-Networks for the Recognition of 3D Point Cloud Models

Published in International Conference on Computer Vision, 2017

The paper presents a novel neural architecture for feature extraction from point clouds based on hierarchical feature pooling from kd-trees constructed on the input point clouds. The architecture is evaluated in several point cloud recognition applications including classification, segmentation and retrieval.

R. Klokov and V. Lempitsky. "Escape from cells: Deep kd-networks for the recognition of 3d point cloud models." In ICCV'17.

Download Paper | Paper Code

Workshop Papers

Large-Scale 3D Shape Reconstruction and Segmentation from ShapeNet Core55

Published in International Conference on Computer Vision Workshop, 2017

The paper presents the results of the ShapeNet point cloud segmentation and single-view reconstruction challenges held at ICCV'17 Workshop. Our segmentation architecture took second place.

L. Yi, et at. "Large-Scale 3D Shape Reconstruction and Segmentation from ShapeNet Core55." In ICCVW'17.

Download Paper