FishServer/FishMeasure/segmentation/README.md

Fish Body Segmentation (YOLOv8-seg)

This folder provides a quick pipeline to train a body-only fish segmentation model using Labelme polygon annotations.

The goal is to produce a mask that excludes fins and tail (as much as possible), so the depth->pointcloud becomes cleaner for weight estimation.

1) Labeling in Labelme

• Use Labelme polygon tool.
• Recommended class name: body (you can use other names; see --classes below).
• Each image produces a .json annotation file.

2) Convert Labelme JSON -> YOLOv8-seg dataset

This will create a YOLO dataset folder:

<out_dir>/
  images/train, images/val, images/test
  labels/train, labels/val, labels/test
  dataset.yaml

Example:

python3 segmentation/prepare_yolo_seg_dataset.py \
  --source_dir /path/to/labelme_export \
  --out_dir ./datasets/fish_body_seg \
  --classes body \
  --train_ratio 0.8 --val_ratio 0.1 --test_ratio 0.1 \
  --seed 42 \
  --copy

Notes:

• YOLOv8-seg label format is: <class_id> x1 y1 x2 y2 ... xn yn (all normalized to [0,1]).
• If an image has no valid polygons, an empty label file will be written (you can change this later if desired).

2b) Filter existing prepared dataset (if only some images are labeled)

If you already have a prepared YOLO-seg dataset but only some images have labels, use this mode to filter and keep only labeled images:

python3 segmentation/prepare_yolo_seg_dataset.py \
  --prepared_dataset /home/ubuntu/data/fish/fish_measure_intermediates/yolo_seg \
  --out_dir ./datasets/fish_body_seg_filtered \
  --classes body \
  --copy

This will:

• Scan images/train/, images/val/, images/test/ for images
• Check for corresponding .txt label files in labels/train/, labels/val/, labels/test/
• Only copy/symlink images that have labels
• Generate a clean dataset.yaml for training

3) Visualize labels (optional - verify conversion correctness)

Before training, you can visualize the converted labels to verify they're correct:

python3 segmentation/visualize_yolo_seg_labels.py \
  --dataset ./datasets/fish_body_seg_filtered \
  --output ./visualizations/yolo_seg_labels \
  --split train \
  --max_images 50 \
  --classes fishbody \
  --alpha 0.5

This will:

• Load images and their corresponding .txt label files
• Draw polygon masks (semi-transparent overlay) on images
• Save visualized images to the output directory
• Useful for checking that Labelme → YOLO conversion preserved polygon shapes correctly

4) Train YOLOv8 segmentation

python3 segmentation/train_yolo_seg.py \
  --data ./datasets/fish_body_seg/dataset.yaml \
  --model yolov8s-seg.pt \
  --epochs 200 \
  --batch 16 \
  --imgsz 640 \
  --project runs/seg \
  --name fish_body_seg_$(date +%Y%m%d_%H%M%S)

Outputs:

• runs/seg/<name>/weights/best.pt

5) Next step (pipeline integration)

After training, you can run segmentation on:

• full image, or
• cropped image from detector bbox (often better when fish is small in the frame).