1,044 scenes containing over 1 million image pairs labeled as either 'doppelgangers' (visually similar but distinct locations) or 'non-doppelgangers' (same physical location). The dataset provides 3D reconstructions and camera poses to serve as ground truth for disambiguating visually identical structures in computer vision tasks.
Use Cases
- Train a binary classifier to filter false positive matches in SfM pipelines using the image pair labels
- Develop image retrieval algorithms that distinguish between visually similar landmarks using the scene-level metadata
- Evaluate the accuracy of local feature descriptors against repetitive structures using the provided 3D ground truth
Strengths
- 1,044 distinct scenes sourced from large-scale internet photo collections
- Binary classification labels for image pairs distinguishing 'true' matches from 'doppelganger' false positives
- Includes camera intrinsic and extrinsic parameters derived from verified SfM pipelines