BoostCompTrack: Dataset for a multi-purpose tracking framework for salmon welfare monitoring

The data relates to code available at  https://github.com/espenbh/BoostCompTrack

• CS_train.zip
  31 labeled images of salmon in cage scenes for training. Annotations include key anatomical features.

• CS_val.zip
  6 labeled images from the same context, used for validation.

• detector_optimization.zip
  Contains 5 Precision–Recall (PR) curves used for evaluating and selecting detection models.

• detector.zip
  Trained YOLOv5 models for salmon detection:

  • keybox_detection: models trained to detect keypoints (e.g., fins, head, body).

  • bounding_box_detection: models trained to detect bounding boxes around salmon.
    Both use consistent label sets.

• TBW_train_1.zip, TBW_train_2.zip, TBW_train_3.zip
  Labeled images (12, 12, and 22 respectively). Annotation content consistent across all. "TBW" refers to a specific scene type or internal naming convention.

• TBW_val.zip
  Video clip used to visualize model tracking performance in TBW scene.

• TS_train.zip
  8 labeled images from TS scenes with annotations.

• TS_val.zip
  Video clip used to visualize model tracking performance in TS scene.

• GH010031_reduced_length.mp4, GH030031.MP4
  Underwater  example video (with camera tilt), used for inspection and testing.

• 20230312_145025_cage15_t0x_yz.avi.avi
  Sample underwater video recorded in cage 15, showing fish behavior under standard conditions. Used for visual inspection and qualitative assessment of model performance.

• 20240508_121121_1715170281739114875_Korsneset_Merd07_22228352_B2Ave_cAIge.avi-0000_00h02m40s_00h00m40s_win.avi
  Cropped segment from stereo footage recorded at Korsneset, cage 07 (camera ID 22228352), showing 40 seconds of activity. Used for model evaluation and behavioral analysis.