keyur-neeved/df40-faces

DF40 face crops

SOURCE SCAN COMPLETE. Review pending: 152,170 rows.

Source: keyur-neeved/df40-filtered. Pinned source revision: 9f96db20325db89c8d4f02ac10fe1f3ded8cf163.

Only face detection and face cropping are applied. There is no width restriction, image deduplication, balancing, blur screening, or split creation. Saved crops retain their native pixel dimensions; PNG encoding is lossless relative to the decoded source crop. Existing source JPEG compression cannot be reversed. The detector internally resizes an analysis copy. EXIF orientation is normalized using integer pixel rearrangements; there is no face alignment or interpolation in saved crops.

Automatic acceptance requires exactly one candidate at at least two SCRFD scales, two high scores, and agreement among every detected rectangle. Any scale with multiple candidates triggers review. Scores and agreement are heuristics, not measured accuracy or a near-100% guarantee. Multiple faces, uncertain detections, and decoding exceptions remain in the review manifests; they are not included in train until a human supplies the correct face rectangle through this notebook. A detected face can still be the wrong subject. Review accepted crops as well. All four original label values are copied unchanged; face detection does not verify them.

The face detector is SCRFD 10G from InsightFace's official v0.7 buffalo_l pack. Only the detection network is used. InsightFace's provided pretrained models have non-commercial research terms, separate from its code license. Original DF40/source dataset terms still apply; this card does not assign a new dataset license.

{
  "scales": [
    128,
    320,
    640
  ],
  "candidate_score": 0.35,
  "accept_score": 0.8,
  "agreement_iou": 0.75,
  "margin": 0.15
}

class_id	class_name	binary_label	binary_name
0	real	0	real
1	FS	1	fake
2	FR	1	fake
3	EFS	1	fake
4	FE	1	fake
5	PA	1	fake

_face_crop/audit/ records every source row, detection decision, rectangle, and output location. _face_crop/review/ retains unresolved source references at the pinned revision. _face_crop/manual/ records any subsequent manual crops. These support files are not extra dataset subsets or splits. Originals stay in the source repository; the review manifests are references, not backup image copies. Each source row appears at most once in train. Similar video frames are all retained if their face crops are accepted. One merged train split does not provide an independent evaluation split; preserve source-video grouping in later evaluation.

{
  "source_scanned": 641902,
  "automatic_crops": 489732,
  "manual_crops": 0,
  "train_rows": 489732,
  "review_pending": 152170
}

from datasets import load_dataset
ds = load_dataset('keyur-neeved/df40-faces', name='image', split='train', token=True)