← All projects
Multimodal AI · Retrieval
Multimodal Geolocation and Reranking
Combining visual embeddings and contextual reasoning to improve location candidates.
ExplorationOverview
An exploration of using vision-language models alongside image encoders for candidate-level reasoning. The goal is to assess whether contextual visual understanding can improve a first-stage geographic retrieval system.
Methods & topics
Highlights
- Investigating VLM-assisted reranking of top-k geographic candidates.
- Comparing first-stage retrieval with refined candidate predictions.
- Studying probability calibration and distance-aware evaluation.