Hi, I'm
Computer Vision Researcher
I build systems that interact with the visual world in 3D — from robust point cloud assembly to scene understanding for robotics and beyond.
I'm a CS Masters student at ETH Zürich focusing on 3D vision research, working at the intersection of deep learning and geometric understanding. My interests span areas including 3D shape assembly, scene registration, and vision for robotics.
TBA
By distilling relational geometric structure from a frozen pretrained 3D encoder into a flow-matching assembly backbone via token-wise cosine and CKA alignment losses, TORA achieves faster convergence (up to 6.9×), improved in-distribution accuracy, and stronger zero-shot transfer across five benchmarks — all with zero additional inference cost.
TBA
By lifting RGB-only egocentric sequences into object-level 3D scene graphs with off-the-shelf foundation models and using a vision-language model to match object instances across sessions, verified via height priors and geometric consensus, PROSE achieves state-of-the-art cross-session registration on egocentric benchmarks without any learned parameters, depth sensors, or training.
To be updated — check back soon!