CraterBench-R: Instance-Level Crater Retrieval for Planetary Scale

A team of researchers has published CraterBench-R, a new benchmark that reframes planetary crater analysis as an instance-level image retrieval problem. The curated dataset contains about 25,000 distinct crater identities with multi-scale images and manually verified queries. The research reveals that self-supervised Vision Transformers (ViTs), especially those pretrained on domain-specific data, significantly outperform larger generic models. A key finding is that using all 196 patch tokens from a ViT for late-interaction matching yields high accuracy, but storing every token for planetary-scale analysis is computationally prohibitive.

To solve this efficiency problem, the team developed 'instance-token aggregation,' a training-free method that clusters similar tokens and aggregates them into a few representative ones. At K=16 clusters, this technique boosts mean Average Precision (mAP) by 17.9 points compared to simple token selection. At K=64, it matches the accuracy of using all 196 tokens but with far less storage overhead. For practical deployment, they propose a two-stage pipeline: a fast single-vector search to create a shortlist, followed by a more accurate reranking using the aggregated tokens. This pipeline recovers 89-94% of the full late-interaction model's accuracy while searching only a tiny fraction of the total dataset, making planetary-scale crater matching and catalog management finally feasible.