lakibrankovic/womrs-yolo

Worm Detection YOLO (8-Channel)

YOLO object detection model trained for worm detection on 8-channel input data.

Model Overview

• Task: Object Detection
• Framework: Ultralytics YOLO
• Input: 8-channel images (not standard RGB)
• Classes:
  • 0: worm
  • 1: worm-under

Important Note

This model is trained on 8-channel data and is not directly compatible with standard 3-channel RGB images without adapting preprocessing.

Training Details

• Model variant: [e.g. yolo11n]
• Image size: 640
• Epochs (planned / actual): [200 / 137]
• Best checkpoint: best.pt
• Best metric (val mAP50-95): [0.5741]
• Dataset split: train / val / test via data.yaml

Files in This Repository

• best.pt — best checkpoint (recommended for inference)
• [optional: last.pt] — last epoch checkpoint
• [optional: results.csv] — training metrics
• [optional: args.yaml] — training args

Intended Inference Pipeline

Inference is expected with a custom multichannel pipeline (example):

• load .npy (8 channels)
• run tiled/sliced inference (SAHI-like approach)
• apply postprocessing (NMS + nested suppression)

If you share this model, also share the preprocessing + inference script to ensure reproducibility.

Example (Ultralytics load)

from ultralytics import YOLO

model = YOLO("best.pt")
# NOTE: Default model.predict(...) assumes standard image input.
# For this model, use your custom 8-channel preprocessing/inference pipeline.