---

name: data-preparation description: "Prepares and validates PaddleDetection COCO, VOC, WIDER, MOT, keypoint, custom, semi-supervised, and sliced-image datasets." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Data Preparation

Use this route for annotation conversion, dataset directory layout, label lists, custom readers, COCO/VOC/MOT/keypoint data, semi-supervised splits, or small-object slicing.

Workflow

1. Decide the task and schema before touching files: detection/instance segmentation, keypoint, MOT, rotated detection, or industrial pipeline input.
2. Choose the matching dataset config under the target checkout's configs/datasets/ and record its dataset root, image directory, annotation path, category/label source, metric, and num_classes.
3. Validate a tiny sample with scripts/validate_detection_dataset.py before downloading full data or starting workers.
4. For custom annotations, prefer COCO JSON or VOC XML. Use the documented converter command patterns in references/conversion-workflows.md; keep the source data and generated annotations separate.
5. For MOT, verify images/, labels_with_ids/, sequence metadata, normalized boxes, class IDs, and identity IDs. For keypoints, verify the keypoint order and visibility convention.
6. For semi-supervised or sliced datasets, make the output directory explicit and preserve the random seed, percent, slice size, overlap, and merge method.

Schema rules

• COCO detection boxes are [x, y, width, height]; image IDs and category IDs must be consistent.
• VOC XML boxes are corner coordinates xmin, ymin, xmax, ymax; label names must match the label list/config.
• MOT labels use [class, identity, x_center, y_center, width, height] normalized to image dimensions; identity=-1 means unknown.
• COCO keypoints use triples [x, y, visibility] in the model's expected order; missing/unlabeled points use the documented zero convention.

Read references/data-formats.md, references/conversion-workflows.md, and references/troubleshooting.md before launching training. Download helpers are network/large-data operations and are not run by the bundled validator.