For disease recognition, microscopy image recognition, object detection and visual classification: organise images, annotate, run QA and split training sets.
Organise images by class, source, date or experimental batch, and standardise names and formats.
Keep the original images; never annotate or compress the originals directly.
Define classes, bounding box rules, segmentation rules and how to handle difficult cases.
Agree on rules before multiple people annotate, or the dataset will be inconsistent.
Annotate for classification, object detection, semantic segmentation or instance segmentation depending on the task.
Annotate a small batch as a trial run before scaling up.
Spot-check for wrong labels, missed objects, duplicate annotations and class imbalance.
Data quality usually affects final performance more than model architecture.
Export data in YOLO, COCO, Pascal VOC or classification-folder formats.
Before exporting, confirm there is no data leakage between training, validation and test sets.
