🚧 Multi-Subject Human Pose & Facial Expression Dataset (Studio, 4K) 🚧

Human Pose & Expression Dataset

A studio-captured dataset of 16 individual human subjects, comprising 10,761 images. Subjects were photographed indoors under controlled studio conditions in Seoul, Republic of Korea, across five capture blocks (snap, zoom, pose, facial expression, and video). Every asset carries structured meta capture category, framing (sub-category), pose angle, facial expression, apparent demographic attributes, technical properties, and a face bounding box.

Motivation

• Purpose. To provide a consistently-captured, richly-annotated set of human portrait, pose, and expression imagery and video for AI/ML training in computer vision (pose/expression/movement recognition).
• Created and funded by. Produced by SpotLite Co., Ltd. (Seoul, Republic of Korea); commissioned and funded by Mozilla Data Collective Limited (London, United Kingdom). Each subject was compensated (KRW 450,000 before withholding tax) per the model participation consent form.

Composition

• Instances. Individual human subjects; each instance is one photograph or one video clip of a single person (people_count = 1).

• Total. 10,761 images across 16 subjects (~638–884 assets per subject).

• Formats. Images: JPEG, 3330 × 5000 px.

• Per-asset metadata (CSV, one row per asset): country, region, asset_id, pose_angle, expression, sub category, model_release_on_file, property_release_on_file, asset_type, category, color, file_size_bytes, dimensions, video_length, video_language, people_count, apparent_gender, apparent_ethnicity, apparent_age_range, face_bounding_box, background_type.

• Capture categories (per asset): snap photo = 3,155; pose photo = 2,730; face expression = 2,550; zoom photo = 2,326; video = 536.

• Framing (sub category): upper body = 4,683; full body = 3,835; half body = 2,384; half body (chair) = 266; full body (sitting) = 94; full body (chair) = 35.

• Pose angles: Frontal (0–15°) = 6,726; Three-Quarter/Half-Profile (15–45°) = 2,359; Profile (75–90°) = 2,151; Rear/Back (180°) = 60; Looking Down = 1.

• Facial expressions: Neutral = 9,042; Happiness = 556; Disgust = 457; Surprise = 376; Fear = 341; Sadness = 263; Anger = 262.

• Apparent demographics (per subject, annotated). Subjects are non-Korean and ethnically mixed; country = KR / region = APAC denote the capture location (Seoul), not subject nationality.

  Model	Gender	Ethnicity	Age range
  01	Male	Black	20–29 (Young Adults)
  02	Male	White	20–29
  03	Female	East Asian	20–29
  04	Female	East Asian	20–29
  05	Male	East Asian	20–29
  06	Female	White	20–29
  07	Male	East Asian	20–29
  08	Female	East Asian	20–29
  09	Female	White	20–29
  10	Female	Mixed Racial	20–29
  11	Male	White	20–29
  12	Male	White	20–29
  13	Female	White	20–29
  14	Male	East Asian	45–64 (Middle-Aged)
  16	Female	White	20–29
  17	Male	White	20–29

Aggregate: 8 male / 8 female; ethnicity: White 8, East Asian 6, Black 1, Mixed Racial 1; age: fifteen subjects 20–29, one subject 45–64. Subject 15 is intentionally not included in this release.

• Face bounding boxes. [x, y, w, h] in pixel coordinates matching each asset's dimensions. Image boxes are present for all subjects. For subjects 06, 07, 08, 09, 11, and 12 the boxes were regenerated by an automated face detector; where a frontal face could not be localized (full-body/profile shots) a capped head-region estimate was used, and video-frame boxes are not provided for these six subjects. Head-region estimates may warrant a spot-check for tight-crop uses.
• Sensitivity. The dataset consists of identifiable images and video of real people, including apparent demographic attributes and biometric features. It contains no subject names in the metadata. Per-subject written consent is on file (model_release_ TRUE).

Collection process

• How. Direct studio capture (high-resolution photography and video) of consenting subjects, indoors under controlled lighting and background, at B1-1, 99 Bomun-ro, Seongbuk-gu, Seoul, Republic of Korea. Each session lasted approximately 1–1.5 hours.
• When. May 2026 (sessions on 5–7 May 2026).
• Consent. A bilingual (English/Korean) Model Participation Consent Form is on file for every subject. Subjects are 18+ and were compensated. The form grants SpotLite an exclusive, perpetual, sublicensable, full-buyout license to use the deliverables (including likeness, image, voice, and biometric features, without the subject's name) for AI/ML training, and to sublicense those rights to Mozilla Data Collective on the same exclusive full-buyout basis. Subjects could withdraw in writing before the shoot; after capture the licenses are irrevocable. Subjects may use their own images for non-commercial personal portfolio only.
• Ethical considerations. No formal external ethics-board review was conducted; the safeguard is per-subject informed consent. Each subject: is 18+ with full legal capacity; signed a written, bilingual (English/Korean) consent form after an opportunity to review it (including optional legal counsel); was compensated; could withdraw without penalty before the shoot; and had the scope of us (AI/ML training only, with the specific prohibitions in Uses) disclosed in the form. No subject names are included in the released data.

Preprocessing / cleaning / labeling

• Media renamed to MDC_model{NN}_image### / _video### and organized one directory per subject.
• Per-asset metadata compiled into a per-subject CSV; fields normalized for cross-subject consistency (asset-id format, asset_type, category, framing/sub category, apparent_gender casing, and video_length format standardized; stray whitespace trimmed).
• Face bounding boxes detected/verified programmatically (see Composition).
• One low-quality clip (subject 17) was removed from both the media and the metadata.
• Media and metadata were reconciled so that every delivered file has exactly one metadata row and every metadata row has exactly one file. Metadata rows without a corresponding delivered image or video were removed (subject 14: 150 images; subject 04: 1 video; subject 16: 1 image), as were non-media working files. After reconciliation the release contains 11,297 assets.

Uses

• Intended use. AI/ML training, development, evaluation, and improvement of computer-vision systems that recognize human poses, facial expressions, and movement.
• Forbidden uses (per the model consent form): public distribution, display, or performance of the deliverables; advertising, social-media campaigns, or any consumer-facing published content; generating deepfakes or AI-synthesized likenesses that impersonate any identifiable individual; and any commercial use outside the AI/ML-training scope (including merchandising, endorsement, or sale of a subject's likeness as a standalone asset).
• Restrictions. Use limited to AI/ML training and research within the consent scope; Subjects' identities and project details are subject to confidentiality obligations.