Based on 22 crops with GT in {cigarette, gun, knife, phone}
Annotated crops: jina_crops/ and nomic_crops/ (conf>=0.5, gap>=0.02)
Raw crops (no label): crops/
Person/car grouping only: detection_crops/

Jina (best accuracy):
  conf_threshold = 0.5
  gap_threshold  = 0.02
  accuracy       = 0.7727

Nomic (best accuracy):
  conf_threshold = 0.5
  gap_threshold  = 0.02
  accuracy       = 0.7273