Rank changes: where fine-tuning helps and where it does not¶
What to look for: the counts are over the whole CzechLynx test split, and the examples are a random sample, so failures are shown as often as they occur.
Pick a query from the CzechLynx test split and compare its top 5 under cosine retrieval, default LoMa and LoMa + WildMatch, with gallery photos of the true individual framed in blue. The filters select queries that fine-tuning rescued, still gets wrong, regressed on, or already had right; the counts above the examples are over the whole test split, and the examples themselves are a random sample of each category.
- Rescued: the fine-tuned matcher ranks the true individual first and the default matcher does not.
- Regressed: the opposite. Both counts matter; the paper reports the net effect.
- Still wrong: neither matcher finds the individual at rank 1. Many of these queries are night or infrared frames, or the individual is outside the candidate list.
- Already right: both matchers rank the true individual first.
All three methods rank the same 250 candidates per query, the paper's main budget. A dash in the true-rank readout means the individual was not among them.
Photographs from CzechLynx (Picek et al.), time-closed split.