Sara Rojas

I'm a PhD Candidate at KAUST in Saudi Arabia, under the supervision of Professor Bernard Ghanem.

Experience in 3D computer vision: neural rendering, 3D reconstruction, 3D-based recognition tasks, diffusion models and robustness.

I've recently completed a research internship at Naver Labs Europe, where I worked on extending MAST3R to better understand humans in-the-wild. I was fortunate to be advised by Gregory Rogez, Matthieu Armando, and Vincent Leroy.

Prior to that, I interned at Adobe Research, where I worked under the guidance of Kalyan Sunkavalli. I also collaborated with Reality Labs at Meta in Zurich, mentored by Albert Pumarola and Ali Thabet. Earlier, I conducted research at the University of Southern California with Autumn Kulaga.

Email  /  CV  /  Scholar  /  X  /  Github

Last updated: April 2025

Research

I’m interested in 3D computer vision, deep learning, generative AI, and image processing. Most of my research focuses on NeRF and its applications, including scene editing and efficiency. Lately, I’ve been working on 3D reconstruction.

We introduce an inpainting approach that leverages the depth information of NeRF scenes to distribute 2D edits across different images, ensuring robustness against errors and resampling challenges.

TrackNeRF enhances NeRF reconstruction under sparse and noisy poses by enforcing global 3D consistency via feature tracks across views, inspired by bundle adjustment. It outperforms prior methods like BARF and SPARF, setting a new benchmark in challenging scenarios.

Re-ReND distills the NeRF by extracting the learned density into a mesh, while the learned color information is factorized into a set of matrices that represent the scene's light field. Re-ReND can achieve over a 2.6-fold increase in rendering speed versus the SOTA.

A neural field representation that integrates a semantic field along with the usual radiance field. SegNeRF inherits from previous works the ability to perform novel view synthesis and 3D reconstruction, and enables 3D part segmentation from a few images.

We perform transferable adversarial attacks on 3D point clouds by utilizing a point cloud autoencoder. We exceed SOTA by up to 40% on transferability and 38% in breaking SOTA 3D defenses on ModelNet40 data.

Kudos to Dr. Jon Barron for sharing his website template.