Integrate Baseline VLM and Hierarchical Models - P6 #1587

Activity

1. kelseylmartin

   added

   Type: Feature RequestNew feature or requestNew feature or request

   on Jan 27, 2026
2. csbrown-noaa commented on Apr 22, 2026

   

   csbrown-noaa

   on Apr 22, 2026

   More actions

   FYI, We have some model training pipelines and weights for hierarchical inference on the GFISHER data already. The major caveat there w.r.t. the UI/UX is that the output prob vector conflicts with standard postprocessing of "flat" predictions - e.g. P("Animalia") is always strictly greater than P("Felis catus") because all Felidae are Animalia by definition, so doing things like "taking the largest probability class" will always select "Biota" and never anything else. Happy to discuss further.
3. PaulHax commented on Aug 21, 2026

   

   PaulHax

   on Aug 21, 2026

   Collaborator

   More actions

   Hi y'all. Some related "higherachrical type" UI work is coming together here:
   #1867

   The type hierarchy, like a domain to species taxonomy, can be defined in an imported kwcoco file:

    "categories": [
       { "id": 1, "name": "fish", "keypoints": ["head", "tail"] },
       { "id": 2, "name": "shark", "supercategory": "fish" },
       { "id": 3, "name": "crab" }
     ],
     "annotations": [
       {
         "id": 1,
         "image_id": 1,
         "category_id": 2,
         "score": 0.97,
         "prob": [0.03, 0.97, 0],
         "dive_confidence_pairs": [["shark", 0.97], ["fish", 0.03]],
         "track_id": 42
       }