Day 52: Fine-tuning YOLO: dataset prep for person + garment classes

Teaching YOLO your garment classes

Pretrained YOLO knows 'person' but not 'saree' or 'kurta'. Fine-tuning adds your garment classes. First, the data: YOLO needs images plus label files, one .txt per image, each line class_id cx cy w h with coordinates normalized 0–1 relative to image size. Getting this format exactly right is where most YOLO training failures actually originate.

YOLO label format — one .txt per image, normalized coordinates

# customer_01.txt  (alongside customer_01.jpg)
# class_id  center_x  center_y  width  height   (all normalized 0-1)
0 0.512 0.480 0.230 0.760      # person
3 0.505 0.350 0.180 0.240      # shirt
1 0.500 0.700 0.160 0.300      # trousers

data.yaml — points YOLO at your splits and names your classes

path: ./garment_dataset
train: images/train
val: images/val
names:
  0: person
  1: trousers
  2: dress
  3: shirt
  4: saree
  5: shoes

Split by person, label consistently

Day 38's rules carry over: split by *individual/photoshoot* so the same person doesn't appear in both train and val (leakage), and define ambiguous garment boundaries once and label them uniformly. Annotation tools (Roboflow, CVAT, Label Studio) export YOLO format directly — use one rather than writing label files by hand.

Key terms

YOLO label format: One text file per image; each line is class_id and normalized center-x, center-y, width, height.
data.yaml: A config file telling Ultralytics where the train/val images are and what the class names are.
Annotation tool: Software (Roboflow, CVAT, Label Studio) for drawing boxes and exporting labels in a chosen format.

Produce YOLO-format labels with correct normalized coordinates
Write a data.yaml naming your person + garment classes
Split by person/shoot to avoid leakage

In a YOLO label file, the bounding box coordinates are expressed how?

Day 52: Fine-tuning YOLO: dataset prep for person + garment classes

Teaching YOLO your garment classes

We use cookies