mrtpk/FishEye

Detecting fish from sequences

Visum2020

• visum-2020
• visum-competition: Fish Detection in Underwater Images

Why is it important?

• Offshore aquaculture
• Environmental studies
• Before building a new offshore field ( i.e. offshore wind farm)
• Before decommissioning an old field

Dataset

• Train

  • 84 sequences of 60 frames each;
  • Some images contain one or more fish;
  • Some images do not contain any fish.

• Test (no access)

  • 30 sequences of 60 frames each
  • Some images contain one or more fish;
  • Some images do not contain any fish;
  • Daily test: 10 sequences of 60 frames each.

• Annotations format

  • sequence; frame; [(x_min , y_min , x_max , y_max ), x_min , y_min , x_max , y_max
  • sequence; frame;

• Predictions format

  • sequence; frame; [(x_min , y_min , x_max , y_max )]; confidence

Evaluation metrics

• IOU
• Precision
• Recall
• Average precision (AP@[0.5:0.95] - averaged over different IoU thresholds

Team FishEye

• Members

  • Arjun S Kumar
  • Gabriella M Mansur
  • Thomas Paul

• Discussion Notes

• Presentation Slides

Algorithm

• Ultralytics yolov5

Challenges

• Find ways to deal with the big variability of visibility conditions underwater;

Strategy

• Implemented augmentation Albumentations