Reproduce “Object-Level Explanations for Image Geolocation Models: a GeoGuessr use-case” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Emilie Durrieu, Christophe Hurter, Philippe Muller, Victor Boutin
Researchers from ENAC, IRIT, ANITI, and CNRS developed an object-centric framework to explain image geolocation models by transforming diffuse attribution maps into discrete, object-like elements. This approach demonstrated that these extracted elements consistently retain more predictive information than random regions, enabling a more interpretable understanding of model decisions.
Key insights
- Attribution maps, despite their often diffuse appearance, contain structured predictive signals that can be successfully decomposed into localized, perceptible visual elements without substantial loss of task-critical information.
- Geolocation model predictions can be attributed to specific, localized visual cues (object-like elements), thereby bridging the gap between abstract pixel-level activations and human-interpretable objects.
- The quality and spatial coverage provided by the segmentation model significantly influence the effectiveness of extracting these predictive object-like regions from attribution maps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.