Best Practices for labeling images

Polygons was chosen over bounding boxes to reduce excess noise that is often associated with bounding boxes, this in turn increases the accuracy of the model.

Best Practices We Follow

  1. Only create polygons for each fish that you can clearly see
  2. When cleaning up image polygons (polylines)
    • “zoom in” on the polygon and make sure that it’s very accurately drawn around fish species body shape (tight to body)
  3. Look for bogus polygons that may have been autogenerated and delete them (i.e.: rocks, seaweed/kelp, animals that are not fish- marine mammals, cephalopods, star fish, etc-, humans, boats, etc.)
  4. Meta data settings
    • If “Fish Art or Fish Drawing or Fish Memes” make image as “Not a real fish (drawing) photos”
      • Turn on “Not a real fish (drawing)” flag under meta data on top level image
    • If image if of a underwater fish, mark underwater in the meta data
  5. If you know the name of the fish, tag it with the species name. Otherwise mark as UNKNOWN species
  6. How to deal with human body parts such as fingers, hands, etc.
    • Draw around fishes head and remainder of body stopping at obscured areas.

Examples:

Holding a fish
Obscured area's of the fish Keep the polylines tight to the body shape

Best Practices for Labeling Fish Traits

Each fish species has a unique set of identifying attributes, and these attributes will be tagged with smaller polygons. Training the AI model to recognize the unique identifying attributes of species we will increase the accuracy of the model.

Best Practices we follow

Examples:

Fish Traits Summary
Head Mouth Eye
Head Form Mouth Position Eye
Caudial Dorsal Anal Fin
Caudal Fin Dorsal Fin Anal Fin
Pectoral Fin
Pectoral Fin Pelvic/Ventral Fin