I'm a Software Engineer at HubSpot in London. I completed my Master's in Computer Science at the Technical University of Munich, and did my thesis on object-centric learning at the Computer Vision Lab at ETH Zurich. I received my Bachelor's in Computer Science from Özyeğin University.
Previously, I interned as a Software Engineer at HubSpot in Berlin, working on search and translations in the Forms team.
How well does a feed-forward Gaussian splatting model trained on synthetic scenes work on real photos? We measure the gap for MVSplat and test DINOv2 features, VGGT depth and depth supervision.
We revisit the Splatter Image framework and show how semantics from vision-language models, depth priors, and tailored loss functions can resolve ambiguity, improving view consistency and geometric accuracy.
While widely used in 3D reconstruction, the effectiveness of training on synthetic data is rarely questioned. Can models trained in synthetic environments truly handle real-world complexity?