Dataset preparation¶

Modern object detection, instance segmentation and oriented object detection¶

Select the dataset root containing data.yaml; the oriented detection page also accepts a YAML path. YAML specifies training, validation and class mapping. Images and labels must correspond; class IDs start at 0 and match the class list.

dataset/
  data.yaml
  images/train/
  images/val/
  labels/train/
  labels/val/

Task	Label contents: one object per line
Standard detection	class_id center_x center_y width height
Instance segmentation	class_id x1 y1 x2 y2 …, an object's contour polygon
Oriented object detection	class_id x1 y1 x2 y2 x3 y3 x4 y4, four corners

Normalize coordinates to 0–1. OBB corners must run clockwise or counterclockwise. Ordinary five-column boxes, raw DOTA labels and self-intersecting quadrilaterals are not valid OBB input. The software checks labels and reports file and line numbers.

Modern classification¶

Organize training/validation images by class, using the corresponding subfolder names as class names. Explicit train/class_name and val/class_name folders are recommended; the classification engine reads them according to its dataset rules. This page does not import classic classification TXT lists.

classification/
  train/OK/...
  train/NG/...
  val/OK/...
  val/NG/...

Classification results are written beside the dataset in dataset-name_runs, preventing results from being treated as a class. Keep old result folders out of class directories.

Classic object detection: one COCO folder¶

COCO mode is selected by default. Select the dataset root and click Check dataset. The software recognizes training images, COCO JSON, classes and an optional independent validation set. A typical layout is:

coco/
  images/train/...
  images/val/...             # optional
  annotations/train.json
  annotations/val.json      # optional

Without an independent validation set, all training images are used for training, with no automatic split or mAP calculation. OK background images may have no boxes: keep their image records without fabricating entries in annotations or creating an OK object class. The software retains these negative samples.

Class mapping is based on the training annotations; training and validation class names must match. The software creates annotation copies and a class list inside the experiment without changing the original COCO files.

Classic classification: one dataset folder¶

Read train_list.txt, val_list.txt and Label.txt first; without lists, detect train/class_name and val/class_name. Preserve the existing split without automatic resplitting. Class IDs must be contiguous, and validation classes must belong to the training classes.

For other folder naming conventions, first run Check dataset and inspect recognized counts and classes. Do not invent class IDs to bypass validation.

Anomaly detection: OK images and independent calibration¶

Include only OK images in the training directory. An independent OK calibration directory is optional; if blank, approximately 20% is held out after content-based deduplication, requiring at least 3 distinct OK images. Calibration images are excluded from the memory bank; content duplicates between training and calibration are rejected.

This automatic holdout rule applies only to anomaly detection, not classic COCO detection.