Vibhas Vats

Work / Internships

Work Experience

Research Intern • Oak Ridge National Laboratory
2025-09 – 2026-03
Knoxville, TN, USA
Impact: Made direct polygon extraction practical at high resolution — ~14× faster inference, ~10× lower memory.

Seven months with ORNL’s GeoAI group on the engineering problem underneath large-scale mapping: turning satellite imagery into clean building and road polygons, across regions large enough to be useful.

Methods that predict polygons directly work well on small tiles with modest vertex budgets, and break down past that. The bottleneck is vertex matching — deciding which vertex connects to which is posed as a permutation over the full vertex budget, so compute and memory grow quadratically. Covering more ground means larger tiles, larger tiles mean denser scenes and more vertices, and the cost grows with the square of that increase. In practice this forces production pipelines to grind through many small, redundant tiles.

I redesigned that stage. Instead of matching over the full budget, the model scores connectivity only across the vertices actually present in each batch, and stops autoregressive decoding when the sequence ends rather than running to a fixed length. The result is roughly 14× faster inference at the same architecture while carrying nearly twice the vertex budget, with memory that stays almost flat as the budget grows — making 512×512 tiles at up to 1200 vertices per image feasible on a single GPU, a setting where the prior approach requires an estimated 278 GiB and simply does not run.

  • Built and deployed an end-to-end pipeline for building and road footprint extraction from satellite imagery, running ~14× faster than the prior state of the art at matched architecture and a higher vertex budget
  • Redesigned the polygon connectivity module to match over valid vertices per batch rather than a fixed budget, reducing permutation cost from O(N²) to O(K²ₘₐₓ)
  • Cut inference memory by roughly an order of magnitude, making high-resolution tiles and dense scenes feasible where they previously were not
  • Added a Fourier-feature coordinate embedding that improved boundary and topology accuracy at no additional matching cost
  • Trained and evaluated across Orbital-Net, INRIA, SpaceNet-2, and Massachusetts Roads
Company
Research Intern • Oak Ridge National Laboratory
2024-05 – 2024-08
Knoxville, TN, USA
Impact: Summer research that became a SIGSPATIAL '25 paper on diffusion-based land-cover forecasting.

Twelve weeks with ORNL’s GeoAI group on a question sitting between generative modeling and Earth science: can a diffusion model forecast how land cover changes over decades?

Most land-use and land-cover forecasting treats the problem as pixel-wise classification or simulation. We asked whether denoising diffusion --- normally used to generate images --- could instead generate plausible futures, conditioned on what a place looked like before and on auxiliary information about it. The target was imperviousness: the paved, built-up fraction of the landscape, which serves as a direct proxy for urban growth and feeds into flood risk, heat exposure, and infrastructure planning.

The summer ended with a model trained across the continental United States on the National Land Cover Database. The work became a SIGSPATIAL ‘25 paper.

  • Developed a conditional denoising diffusion model to forecast decadal land-cover imperviousness change across the continental US
  • Designed a multi-branch conditional normalization scheme, allowing the model to condition on historical land cover and auxiliary covariates simultaneously
  • Trained at scale on distributed GPU clusters (8×A100) using the National Land Cover Database
  • Collaborated with a multidisciplinary team spanning geography, urban science, and machine learning
Company