CIS 5190 · Applied Machine Learning

Image-to-GPS Localization Pipeline

End-to-end image geolocation pipeline fusing CNNs and transformer models. Ranked 1st on the course leaderboard.

Period

Nov – Dec 2025

Tags

Deep Learning · CNN · Transformers · Python

Highlights

  • 01Fusion architecture combining CNN feature extractors with transformer encoders
  • 02~5 m localization error on self-collected validation data
  • 03~13 m error on held-out test data
  • 041st place out of all course teams on the final leaderboard

Write-up

We treated geolocation as a regression problem over visual features, training a hybrid CNN-transformer model on a curated dataset of campus imagery. The transformer head allowed the network to attend over spatially distributed cues — architectural details, sightlines, ground texture — that pure convolutional backbones tended to under-weight.

Links