Fadi Khatib

I am a Ph.D. candidate (direct-track) in the Department of Computer Science and Applied Mathematics at the Weizmann Institute of Science under the supervision of Prof. Ronen Basri.

I completed my M.Sc. at the Weizmann Institute of Science. Prior to that, I received my B.Sc. in Computer Science at the age of 18 from the University of Haifa through the Challenge program, which is designed for exceptional high school students.

My research focuses on deep learning for 3D vision, particularly Structure-from-Motion, camera pose estimation, and visual localization. I have also worked on Vision-Language Models (VLMs) and Vision-Language-Action models (VLAs).

Email  /  CV  /  Google Scholar  /  X  /  LinkedIn

• [10/2026] Our paper What Questions Can Internal Activations of Visual Foundation Models Answer? was accepted to WACV 2027.
• [09/2026] I will give a talk, Learning and Geometry for Structure-from-Motion, at the Workshop on Geometry, Learning and 3D Understanding 2026 at Lund University.
• [09/2026] I will present GSVisLoc at the NeuSLAM Workshop and VGPA at the Wild3D Workshop at ECCV 2026.
• [06/2026] Our paper Learning Global Camera Poses from Noisy View-Graphs for Structure from Motion was accepted to ECCV 2026.
• [06/2026] I will attend the International Computer Vision Summer School (ICVSS 2026) in Sicily.
• [08/2025] Our paper GSVisLoc: Generalizable Visual Localization for Gaussian Splatting Scene Representations was accepted to the CALIPOSE Workshop at ICCV 2025.
• [02/2025] Our paper RESFM: Robust Equivariant Multiview Structure from Motion was accepted to ICLR 2025.
• [10/2024] Our paper Consensus Learning with Deep Sets for Essential Matrix Estimation was accepted to NeurIPS 2024.
• [2024] Awarded the Council for Higher Education Fellowship for Outstanding Doctoral Students, 2024-2027.

* denotes equal contribution

Coming soon...

We present VGPA, a deep global Structure-from-Motion framework based on learned view-graph aggregation. A permutation-equivariant, edge-conditioned graph neural network takes noisy pairwise relative poses as input and outputs globally consistent camera extrinsics, trained without ground-truth supervision using only a relative-pose consistency objective. Followed by triangulation and robust bundle adjustment, the approach is efficient, scales to more than a thousand images, and is robust to graph density.

We present GSVisLoc, a generalizable visual localization method for 3D Gaussian Splatting (3DGS) scene representations. Our approach estimates camera poses by aligning features between a query image and a 3DGS model, followed by refinement for improved accuracy. GSVisLoc achieves strong results on standard benchmarks, outperforms existing 3DGS-based baselines, and generalizes well to unseen scenes without additional training.

We present a robust solver based on permutation equivariant architecture for the structure from motion problem. Our proposed method can cope with a point track tensor contaminated with many outliers. In addition, we modified the bundle adjustment module to make it robust enough to handle classification errors. Our method achieves high accuracy while reducing the time significantly compared to classical methods.

We present NACNet, a Noise-Aware Deep Sets framework to estimate relative camera pose, given a set of putative matches extracted from two views of a scene. We demonstrate that a position denoising of inliers and noise-free pretraining enable accurate estimation of the essential matrix. Our experiments indicate that our method can handle large numbers of outliers and achieve accurate pose estimation superior to previous methods.

We propose to leverage Image Matching (IM) as a pre-trained task for relative pose regression. Specifically, we use LoFTR, an architecture that utilizes an attention-based network pre-trained on Scannet, to extract semi-dense feature maps, which are then warped and fed into a pose regression network.

Developed a method for hands gestures detection, to be used for emotion recognition problem. By utilizing pre-trained models from Google Mediapipe for pose and hand pose estimation.

Worked on mapping human viral proteins, where the goal is to reveal the relation between different viruses by using computational biology methods and applying them to biological databases.

• [09/2026] Learning and Geometry for Structure-from-Motion, Workshop on Geometry, Learning and 3D Understanding, Lund University, Sweden.
• [07/2026] Toward End-to-End Deep Learning for Accurate, Large-Scale Structure-from-Motion, Co-AI Research Group, University of Catania, Italy.
• [01/2026] Toward End-to-End Deep Learning for Accurate, Large-Scale Structure-from-Motion, NVIDIA Research Israel.

• Initiating various initiatives for the Arab students at the Weizmann Institute of Science:
  • Organizing outreach programs, panels, mentoring sessions, social activities, and many others.
  • Supported by the Diversity and Inclusion Office at the Weizmann Institute.

• The Council for Higher Education Fellowship for Outstanding Doctoral Students, 2024-2027.
• The Cornell, Maryland, Max Planck Pre-doctoral Research School, 2023.
• Yoel Meches Memorial Fellowship for Excellent Students, Weizmann Institute of Science, 2023.
• The Feinberg Graduate School Diversity Appreciation Certificate for Advancing Equality, Diversity, and Inclusion, Weizmann Institute of Science, 2023.
• Moshe Horovitz Memorial Fellowship for Excellent Students, Weizmann Institute of Science, 2022.
• Summa Cum Laude (B.Sc Computer Science), University of Haifa, 2022.
• The Council for Higher Education Scholarship for Outstanding Master Students, 2021-2023.
• Young Weizmann Scholars Diversity and Excellence Program Scholarship, 2021.
• Akavia Scholarship for Outstanding Students at the Department of CS, University of Haifa, 2021.
• Dean Scholarship (Summa Cum Laude), University of Haifa, 2020.
• Akavia Scholarship for Outstanding Students at the Department of CS, University of Haifa, 2020.
• The 100 Most Outstanding Students Award at the National Mathematics Olympics, 2018.
• First Place at the "Moofet" National Mathematics Olympics, 2015.

This website is based on Jon Barron's template (source code)