I’m a Senior Applied Scientist at Apple where I work on Apple Intelligence, Visual Intelligence, and contribute to the Apple Foundation Model (AFM). My work centers on multimodal foundation models and their deployment in real-world applications, with expertise in 3D and multimodal scene understanding. I build advanced AI systems that integrate vision, language, and spatial reasoning.
Previously, I addressed the challenge of limited labeled 3D data by developing weakly and self-supervised learning pipelines for images, point clouds, and geometric meshes, resulting in several publications. Before joining Apple, I was a Senior Research Scientist at Fujitsu Research of Europe.
I hold a Ph.D. in 3D computer vision from UCL, London, supervised by Prof. Jan Boehm, with close collaboration with Prof. Tobias Ritschel. In summer 2021, I interned at Adobe in the Creative Intelligence Lab, working on geometrically-driven single-image relighting under Dr. Julien Philip.
arXiv 2026
A video tokenizer that represents videos as variable-length coarse-to-fine token sequences, enabling more efficient generative modeling. VideoFlexTok achieves comparable generation quality with a 5x smaller model and generates 10-second videos using 8x fewer tokens than 3D grid tokenizers.
CVPR 2026
We conduct a comprehensive study of visual structural output capabilities for MLLMs with SO-Bench, covering four visual domains with over 6.5K JSON schemas and 1.8K curated image-schema pairs, revealing persistent gaps in schema-compliant outputs.
ICCV 2025
Building on the Cubify Anything CA-1M, we generate VQA question pairs using an automated pipeline. We show reformatting high-quality 3D data in this way allows us to achieve SoTA results on many 3D spatial reasoning benchmarks.
CVPR 2025
We scale 3D object detection to every object in indoor scenes. Our work demonstrates that as we scale to smaller objects, 3D inductive priors become less valuable and a fully-transformer architecture out-performs SOTA 3D networks.
NeurIPS 2024
A framework for training any-to-any multimodal foundation models. Scalable. Open-sourced. Across tens of modalities and tasks.
EuroGraphics 2022
We address the problem of single image relighting. Our work shows monocular depth estimators can provide sufficient geometry when combined with our novel 3D shadow map prediction module.
International Conference on 3D Vision (3DV) 2021
A novel method for self-supervised monocular 3D object detection. This is achieved through differentiable rendering and a GAN-like critic loss.
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci 2021
A pipeline which demonstrates that Terrestrial Laser Scanning (TLS) 3D data can be automatically labelled using off-the-shelf 2D semantic segmentation networks. With only a simple projection of a panoramic image, strong results can be generated with no additional training.
European Conference on Computer Vision (ECCV) 2020
We present a novel weakly-supervised approach for 3D object detection. Our method can be trained on upto 95% less labeled data and still benefits from unlabeled data.
ISPRS Journal of Photogrammetry and Remote Sensing 2019
Manually labelling buildings for segmentation is a time consuming task. We show that readily available GIS mapping data can be used as training data. We develop a novel pipeline which uses Active Contours to improve coarse polygons into fine per-pixel label maps.
arXiv preprint 2019
We release a synthetic Mobile Laser Scanning (MLS) point cloud named SynthCity. Every point has a per-class and per-instance classification, along with colour, return intensity, end-of-line indicator and time.
Remote Sensing 2019
A comprehensive review paper on deep learning for 3D sensed data classification.
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci 2019
A key issue when training deep neural networks for outdoor point clouds is the inevitable large data imbalance. For example, a typical street scene will contain orders of magnitudes more ground points than street furniture. We develop a novel solution to apply a weighted augmentation to reduce the class-imbalance.
Progress in Physical Geography: Earth and Environment 2018
Linear topologies can be challenging terrains for SfM pipelines. A key source of error is caused by intrinsic camera distortions. We demonstrate through effective camera pre-calibration, distortions can be significantly reduced.
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci 2018
An experimental assessment addressing the ability to train a deep CNN-based object detector (RetinaNet / Faster R-CNN on a low quantity of training data. Specifically in the context of repetitive features (railway track).
JP, US US20230298335A1 • Fujitsu Ltd • Filed: January 25, 2023
This patent presents a novel computer-implemented method for training object detection models without requiring manual annotations. The approach leverages embedding techniques and unsupervised learning to automatically identify and classify objects in images, significantly reducing the time and cost associated with traditional supervised training methods.
US US11972512B2 • Adobe Inc. • Filed: January 25, 2022
This patent describes methods for directional editing of digital images using machine learning techniques. The system enables users to make targeted modifications to specific regions or aspects of images while preserving overall image quality and coherence through advanced neural network architectures.
