Fish monitoring using YOLO11 Object Detection model + BotSort Tracking technique + Customized Counting algorithm
- This Project focuses on river herring detection, all other type of fishes are considered as non-herring class.
- The main contribution includes:
- The curated high-quality balanced training dataset (12,000 images)
- The trained YOLO11 Object Detection Model weights on our dataset
- The customized counting algorithm leveraging accumulated fish detections of 70% of the frames.
git clone https://github.com/Sea-queue/Fish-detection-and-counting.git.git from ultralytics import YOLO
# Load a model with our trained weights
model = YOLO("path_to_the_weights") # load a custom modelThe project includes an evaluation framework that benchmarks three counting algorithms (original, zone-based, and stitch/ReID) across multiple test datasets.
For detailed testing instructions, dataset descriptions, CLI options, HPC setup, and diagnostic output — see test_scripts/TESTING.md.
Quick start:
# Run all 3 counting algorithms on all test videos
python eval.py count \
-d count5 nemasket_normal nemasket_huge nemasket_extra_huge non_herring_easy \
-w model/train113_weights.pt \
--algorithm all \
--no-display \
--device cuda