Just Zoom In: Cross-View Geo-Localization via Autoregressive Zooming

4 Steps

1. 1

   Perform Initial Coarse Localization: Begin with a traditional cross-view geo-localization (CVGL) method. Use it to obtain an initial, broad estimate of the location (latitude, longitude) and a corresponding coarse zoom level (e.g., zoom 15-16). This serves as the starting point for iterative refinement.

2. 2

   Generate Multi-Scale Overhead Context: Based on the current estimated location, programmatically fetch or render a set of overlapping overhead image patches. For the next iteration, generate these patches at a *finer* zoom level (e.g., zoom 17-18) to progressively 'zoom in' on the area of interest. Ensure sufficient coverage for the next search.

3. 3

   Extract Embeddings for Matching: Utilize your pre-trained contrastive embedding model. Extract feature vectors (embeddings) from the input street-view image and from each of the newly generated multi-scale overhead image patches. The model should be robust to minor scale and perspective variations within a local context.

4. 4

   Match, Refine, and Iterate: Compare the street-view image embedding with the embeddings of all overhead patches to find the highest similarity score. Identify the best-matching overhead patch or region. Update your estimated location based on this match. Repeat steps 2-4, iteratively zooming in and refining the location until a desired precision is achieved or a maximum number of iterations is reached.