Ramana Sundararaman

Deformation Recovery: Localized Learning for Detail-Preserving Deformations

Authors: Ramana Sundararaman, Nicolas Donati, Simone Melzi, Etienne Corman, Maks Ovsjanikov

LJN is lightweight method for detail-preserving shape deformations using local Jacobians, where each triangle is considered as training example instead of the entire shape. LJN is thus data-friendly and can learn high-quality deformation (human) from as few as 60 pairs of shapes.

Presented at: SIGGRAPH-Asia, Tokyo, 2024

Physical Property Estimation and Optimization via Constrained Latent Space Exploration

A 3D Generative approach that estimates and improves physical properties while ensuring geometric plausibility by using latent space sampling within convex polytopes. Introduces a new dataset of annotated bottles.

Authors: Ramana Sundararaman, Jiqiong Qiu, Romain Savajano, Matthieu Pichaud, Maks Ovsjanikov

Under-Review, 2024.

Self-Supervised Dual Contouring

SDC introduces self-supervised dual contouring for isosurface extraction, replacing supervised training with novel losses enforcing mesh-SDF consistency. It improves mesh extraction from SDFs that are produced by Deep Networks.

Authors: Ramana Sundararaman, Roman Klokov, Maks Ovsjanikov

Presented at: CVPR (Spotlight), Seattle, 2024

Reduced representation of deformation fields for effective non-rigid shape matching

Marrying mesh-free approximation method with MLPs for non-rigid shape correspondence. Learning reduced deformation parameters to reconstruct smooth deformation, our approach enables efficient, limited supervision and achieves state-of-the-art results on shape matching benchmarks.

Authors: Ramana Sundararaman, Riccardo Marin, Emanuele Rodola, Maks Ovsjanikov

Presented at: NeurIPS, New Orleans, 2022

Implicit field supervision for robust non-rigid shape matching.

An auto-decoder framework which learns continuous deformation fields for shape deformation, using Signed Distance Functions (SDFs) as regularization. Achieves strong performance on noisy, real-world data despite training on clean meshes.

Authors: Ramana Sundararaman, Gautam Pai, Maks Ovsjanikov

Presented at: ECCV (Oral), Tel-Aviv, 2022

Tracking pedestrian heads in dense crowd.

We introduce CroHD, a large annotated dataset for tracking in dense crowds, along with IDEucl metric to evaluate identity preservation. We also propose HeadHunter, a head detector combined with particle-filter-based tracking framework, which acheives superior results compared to state-of-the-art pedestrian trackers.

Authors: Ramana Sundararaman, Cedric Braga, Eric Marchand, Julien Pettre

Presented at: CVPR, (Virtual), 2021

Solving Inverse Computational Imaging Problems using Deep Pixel-level Prior.

Uses autoregressive models as flexible signal priors for inverse imaging problems, enabling better reconstruction of images with texture.

Authors: Akshat Dave, Anil Vadathya, Ramana Sundararaman, Rahul Baburajan, Kaushik Mitra

Accepted at: Transactions on Computational Imaging, 2018

.

Implicit latent space exploration for Shape Optimisation and Correspondence.

Master's Thesis, 2020