Address robustness to degradation and sparse historical samples

Improve robustness of the goldsmith’s-mark retrieval pipeline to physical degradation and to the long-tailed distribution of historical samples, including target groups containing only two marks.

Background

The feature-space analysis shows that some artist-specific clusters remain overlapping even after metric-learning fine-tuning. The paper attributes this unresolved difficulty to the physical degradation of historical marks and the long-tailed sample distribution, in which many target groups contain very few examples. These factors remain barriers to reliable fine-grained retrieval and are explicitly identified as open challenges.