Skip to content

dataset_inspect

Comprehensive dataset statistics and quality intelligence for YOLO or COCO exports.

YOLO and COCO adapters populate the same internal DatasetInventory; shared analysis then computes class balance, annotation counts, bounding-box anomalies, perceptual duplicates, split leakage, quality score, and auto-fix guidance once. Public output fields remain backward compatible.

Inputs

  • dataset_path: dataset root
  • format: auto, yolo, coco, or voc (voc detection is accepted but inspection remains unsupported)
  • include_samples: include sample image paths
  • max_images_analyzed: cap expensive image decoding and fingerprinting for large datasets

The complete inventory still reports all declared/discovered images and annotations. The analysis cap limits only image decoding, dimensions, perceptual hashes, and sample-path evidence.

Outputs

  • detected format, image/annotation counts, normalized class statistics, and orphan counts
  • split detection, image/bounding-box/count histograms, warnings, and sample paths
  • quality_score, class_balance_gini, label_anomalies, duplicate groups, and leakage findings
  • human-readable dataset_card and non-destructive auto_fix_plan

COCO pixel boxes are normalized to image-relative coordinates for shared anomaly checks while raw pixel areas remain available for the existing COCO histogram semantics.

Examples

{"dataset_path":"~/data/mini_yolo"}
{"dataset_path":"~/data/retail","format":"yolo","include_samples":false}
{"dataset_path":"~/exports/coco128","format":"coco","max_images_analyzed":5000}

Common Errors

  • FOVUX_DATASET_001: missing dataset path
  • FOVUX_DATASET_002: unsupported or malformed dataset format
  • FOVUX_DATASET_003: zero images discovered

dataset_validate, dataset_find_duplicates, dataset_split