Research
My research asks a simple question: how can machines understand human motion when cameras alone aren’t enough? Answering it has taken me across computer vision, 3D human modelling, differentiable simulation, and radar signal processing — usually all at once.
A note on what’s listed here. Several first-author manuscripts from this work are currently under double-blind review, so the descriptions below stay deliberately high-level. The published work is linked in full.
Temporally coherent 3D humans from video
Most human mesh recovery models treat video as a stack of independent frames, producing jittery, implausible motion. I designed a structured transformer decoder that recovers the full expressive body — body, hands, and face — from monocular video, with temporal coherence built in as a first-class objective rather than a post-processing step.
The system fine-tunes a ViT-Huge-scale vision foundation backbone parameter-efficiently (LoRA) with mixed-precision training and model distillation, and is evaluated across 12 datasets totalling ~12 million samples, achieving state-of-the-art whole-body accuracy on standard benchmarks. First-author manuscript under review.
Differentiable radar simulation for human sensing
Radar is a compelling sensor for human activity understanding — it works in darkness, through occlusion, and without capturing identifiable images. But labelled radar data is scarce and expensive to collect. I build differentiable, mesh-faithful simulators that render radar micro-Doppler directly from animated 3D human body meshes, distilled from expensive physical-optics renders across seven motion datasets and 60/77 GHz radar platforms.
Synthetic data from these simulators improves real-world zero-shot recognition by +7.4 pp overall and +8.0 pp in an unseen room — without a single new radar capture. First-author manuscript under review.
Do vision foundation models transfer beyond vision?
Spectrograms aren’t natural images, yet the question of whether frozen vision foundation models can read them was surprisingly open. I benchmarked frozen DINOv2, CLIP, and MAE features on radar and WiFi micro-Doppler recognition under cross-sensor shift, against CNN, JEPA, and generative adaptation baselines.
The headline: a frozen DINOv2 probe more than doubles cross-sensor recognition over the strongest matched CNN, while extra synthetic augmentation doesn’t move frozen-feature performance — evidence that cross-sensor recognition is first and foremost a representation transfer problem. First-author manuscript under review.
Radar-based neuroimaging (published)
Can radar see inside the head? With colleagues at Southampton I studied the feasibility of radar for non-invasive brain imaging and tumour detection, simulating electromagnetic interactions in anatomical brain-tissue models (Ansys HFSS) and evaluating signal penetration, fidelity, and safety across Patch and Vivaldi antenna configurations.
The work was published at IEEE EMBC 2024 (paper), presented as an oral talk in Orlando, and received the IEEE EMBC NextGen Scholar Award.
Earlier research
- Universal Gene Transformer (MSc thesis, Distinction) — a 1D-CNN + self-attention classifier over 60,660-gene TCGA RNA-seq profiles spanning 33 cancer subtypes, reaching 92–94% accuracy and matching or beating published baselines without manual feature selection; an explainability pipeline converted saliency maps into gene-effect scores that recovered known biomarkers via TCGA cross-reference and MSigDB enrichment.
- Cooperative swarm robotics for disaster response — a heterogeneous swarm of Thymio ground robots and Tello drones with a virtual mesh network, delivered as a physical demonstrator with DSTL as industry sponsor.
- Self-supervised motion transfer — a faithful reimplementation of a latent image animator that animates still images from driving video, reproducing the full training and evaluation pipeline.
