yasdasd123/easyportrait-lips-open

EasyPortrait lips open, teeth visible (before only)

Filtered subset of EasyPortrait (Kapitanov, Kvanchiani, Kirillova, 2023), taken from the private visible-teeth set.

These are before portraits, mostly webcam framing, where the lips are open enough to show a real smile and the opening is filled by teeth. Barely parted lips, shouts, and tongue-out poses were removed. Hollywood, youthful, and natural smile-design styles are applied downstream and are not labels here.

Filter

Started from 1778 portraits that already had teeth_px >= 400 and teeth_px / (lips_px + teeth_px) >= 0.15.

On the official parsing mask (class 7 = lips, class 8 = teeth), the opening is the non-lip pixels between the first and last lip pixel of each column. A portrait is kept when:

• mean opening height / lip width is between 0.14 and 0.42
• the opening spans at least 45% of the lip width
• teeth fill at least 45% of the opening, and at least 65% when the opening is wide (aspect >= 0.32)
• the tallest column is at most 2.2 times the mean opening, which drops irregular tongue spikes

Scanned: 1778 Kept: 542 Guessed EasyPortrait version: v1

aspect is the mean opening height divided by the lip width. It is not teeth_ratio.

Files

• images/ — RGB portraits that passed the filter
• manifest.parquet — source columns plus between_px, teeth_between, gap_cols, lip_cols, mean_gap, max_gap, aspect, teeth_fill, gap_frac, rel_gap, spike
• filter_config.json — thresholds used for this export

Parsing masks are not included.

License and citation

EasyPortrait is released under a variant of Creative Commons Attribution-ShareAlike 4.0. This derivative stays under the same terms. The SPDX field above is the closest standard id; see the upstream license.

@article{EasyPortrait,
  title={EasyPortrait - Face Parsing and Portrait Segmentation Dataset},
  author={Kapitanov, Alexander and Kvanchiani, Karina and Kirillova Sofia},
  journal={arXiv preprint arXiv:2304.13509},
  year={2023}
}