Pelmeshek/raf-db-7emotions-mediapipe-768

RAF-DB 7 Emotions MediaPipe 768

This dataset is a processed derivative of rhavill/raf-db-7emotions for landmark-based facial expression recognition experiments.

It keeps the source images and labels, remaps labels into a stable seven-class order, creates stratified train, val, and test splits with seed 42, and adds MediaPipe Face Landmarker outputs extracted after resizing each image to 768 x 768 for processing.

Labels

Target label order:

['anger', 'disgust', 'fear', 'happiness', 'sadness', 'surprise', 'neutral']

The label column is a ClassLabel using that order. The label_name column stores the readable class name.

Splits

The source Hugging Face mirror has a single split. This dataset creates deterministic stratified splits:

• —train: 70%
• —val: 15%
• —test: 15%
• —seed: 42

Columns

• —image: source image
• —sample_id: deterministic row id
• —source_dataset: original dataset id
• —source_index: row index in the source dataset
• —split: split name
• —hf_label_id, hf_label_name: original Hugging Face mirror label
• —label, label_name: remapped target label
• —mediapipe_image_size: processing size, fixed at 768
• —landmark_success: whether MediaPipe produced landmarks
• —failure_reason: extraction failure reason if any
• —landmarks_mediapipe_xyz: 478 normalized MediaPipe xyz landmarks
• —landmarks_pixel_xyz_768: 478 xyz landmarks scaled to the 768 x 768 processing frame
• —landmarks_stable_eye_norm: 478 xyz landmarks normalized by stable eye-corner center, scale, and roll
• —bbox_xyxy_768: landmark-derived face bounding box in the processing frame
• —inter_eye_scale: eye-corner scale used for normalization
• —roll_radians: face roll angle used for normalization
• —facial_transformation_matrix: flattened MediaPipe facial transformation matrix when available

Intended Use

This dataset is intended for privacy-conscious facial expression recognition experiments where models consume landmark-derived geometry rather than raw texture-heavy image features.

Caveats

This is a processed Hugging Face mirror dataset. Results should be reported as mirror/proxy results and should include the split seed, MediaPipe processing size, and landmark coverage.

MediaPipe extraction can fail and may fail non-uniformly across classes. Use landmark_success and failure_reason when training or reporting coverage-adjusted metrics.