Summary
When tracking (#1469) merges several detections into one occurrence, the occurrence's determination is chosen by the existing rule: the single highest-scoring classification among all its detections wins. With many detections per occurrence, one outlier decides the label. For example, one "Not identifiable" classification from the small size filter at score 1.0 relabels an occurrence of thirty detections that a classifier labelled consistently.
Directions to discuss
- A vote across the occurrence's detections, weighted by score (for example the mean score per taxon, or the sum of scores), with a minimum number of supporting detections.
- Treat post-processing labels such as "Not identifiable" separately from species labels, so they only win when most detections carry them.
- Keep the current rule for single-detection occurrences, where it is already a vote of one.
Whatever rule is chosen has to apply wherever the determination is computed (update_occurrence_determination and the occurrence's best prediction), not only after tracking, and human identifications keep priority as today.
Related: #1469 (where tracking's results record the determination before and after each run).
Summary
When tracking (#1469) merges several detections into one occurrence, the occurrence's determination is chosen by the existing rule: the single highest-scoring classification among all its detections wins. With many detections per occurrence, one outlier decides the label. For example, one "Not identifiable" classification from the small size filter at score 1.0 relabels an occurrence of thirty detections that a classifier labelled consistently.
Directions to discuss
Whatever rule is chosen has to apply wherever the determination is computed (
update_occurrence_determinationand the occurrence's best prediction), not only after tracking, and human identifications keep priority as today.Related: #1469 (where tracking's results record the determination before and after each run).