Publications

A pipeline for parametric sharp features wireframe extraction from densely sampled point clouds.

A learnable attention-based point neighborhood selection for prediction of the differential geometric properties.

A novel data-driven mesh deformation framework that fits aligned 3D CAD models from a shape database to noisy and partial 3D scans.

We present a new method for vectorization of technical line drawings which consists of (1) a deep learning-based cleaning stage, (2) a transformer-based network to estimate vector primitives, and (3) an optimization procedure to obtain the final primitive configurations.

Struggle to run DensePose network on a mobile device? Check out our work on how to make that possible.

In this work we tackle the video generation problem. Given several first frames, we predict the continuation of a video.

We employ a latent-space Laplacian pyramid representation within a hierarchical generative model for 3D point clouds. We combine recent latent-space GAN and Laplacian GAN to form a multi-scale model for generation of 3D point clouds with gradually increasing levels of detail.

We address the problem of depth map super-resolution with the focus on visual quality of the corresponding 3D geometry. We demonstrate that basing the loss function on deviation of 3D surface rendering instead of direct depth deviation yields significantly improved results as measured by a number of perceptual metrics.

Reconstruction of directional fields is a need in many geometry processing tasks, such as image tracing, extraction of 3D geometric features, and finding principal surface directions. We propose a deep learning-based approach and study the expressive power and generalization ability.

We propose a simple and efficient method for creating realistic targeted synthetic datasets in the remote sensing domain, leveraging the opportunities offered by game development engines. Our evaluations demonstrate that our pipeline helps to improve the performance and convergence of deep learning models when the amount of real-world data is severely limited.

Typical semantic segmentation losses (IoU, cross-entropy) are not sensitive enough to some misalignment of boundaries. As segment is fully explained with its boundary, we propose to use differentiable surrogate of metric BF1 to better account pixels on the edge of a segment.

We introduce ABC-Dataset, a collection of one million Computer-Aided Design (CAD) models for research of geometric deep learning methods and applications. Each model is a collection of explicitly parametrized curves and surfaces, providing ground truth for differential quantities, patch segmentation, geometric feature detection, and shape reconstruction.

We augment a conventional object detection framework with a keypoint detection module and a multi-view consistency loss to make it a robust 3D keypoint estimator, that we use for predicting 3D objects in KITTI road scenes.

We propose a novel multi-texture synthesis model based on generative adversarial networks (GANs) with a user-controllable mechanism. The user control ability allows us to explicitly specify the texture which should be generated by the model. This property follows from using an encoder part which learns a latent representation for each texture from the dataset.

We propose a novel deep learning architecture for three-dimensional porous media reconstruction from two-dimensional slices. Given central slices, we recover the three-dimensional structure around such slices as the most probable one.