PhD candidate in Computer Vision · Universitat de Barcelona
I am a PhD candidate at the Faculty of Mathematics and Computer Science, Universitat de Barcelona, and a member of the AIBA research group. My thesis applies neural radiance fields and 3D reconstruction to realistic, accurate volume estimation of objects, with food and dietary assessment as the driving application. From October 2025 to January 2026 I was a visiting researcher with Benjamin Busam at the Technical University of Munich.
Before the PhD I spent more than ten years as a software engineer (Atypon, Glovo), working in Java, Python, front-end and DevOps. I hold an Honours MSc in Computer Science from the University of Jordan (2020) and an Honours BSc in Computer Science from Zarqa University (2012).
I work on computer vision, deep learning and computer graphics: neural rendering, 3D reconstruction, video segmentation and volume estimation. My PhD is funded by an FPI grant (PREP2022-000101) within the IDEATE project (PID2022-141566NB-I00). My supervisors are:
arXiv 2026 project page / arXiv / code
A cost-normalized benchmark of stacking, substitution and interference: one hand-crafted spectral prior against data-driven alternatives under a single frozen recipe, across 14 datasets and 11 backbones. The free prior is worth more to a small transformer than two rounds of self-supervised pre-training, and nothing to a convolutional network that already has enough data.
arXiv 2026 project page / arXiv / code
A multi-view food video segmentation dataset and benchmark: 55 dish scenes with 25,284 annotated frames under free 360° camera motion, and 20 state-of-the-art segmentation models evaluated alone and combined with video-memory modules.
CVPR 2024 MetaFood Workshop Challenge winner project page / arXiv / code
VolETA estimates food volume from one or a few RGBD images by reconstructing a scaled 3D mesh of the food object, reaching 10.97% MAPE on the MTF dataset.
arXiv 2024 project page / arXiv
FoodMem segments food items from video sequences of 360-degree unbounded scenes. It generates consistent masks of food portions across a video, overcoming the flickering and prohibitive inference speeds of existing semantic segmentation models.
arXiv 2024 project page / arXiv
Orthogonal moments inside convolutional networks for one- and few-shot neural rendering: fewer artefacts, lower learning complexity and less noise than earlier few-shot NeRFs.
arXiv 2023 project page / arXiv
Overcomes several obstacles of earlier NeRFs on real 360° captures: handling multiple video inputs, selecting keyframes, and extracting poses from ambiguous, symmetrical real-world frames.
Theses at the Universitat de Barcelona, co-supervised with Petia Radeva and, depending on the thesis, Ricardo Marques or Albert Clop.
Inter-university Master's Degree in Health Data Science (MHEDAS).