Datasetpublic

notgoodkeeper/cnn-based-drowsiness-detection-data

CNN-Based Drowsiness Detection - Dataset Preprocessed, auto-labeled face-crop images used to train the model in notgoodkeeper/cnn-based-drowsiness-detection. Code: https://github.com/not-good-keeper/cnn-based-drowsiness-detection Collection Frames were captured from a webcam, then run through: Haar Cascade face detection -> crop + pad + resize to 412x412 MediaPipe Selfie Segmentation -> background replaced with white CLAHE contrast normalization -> grayscale… See the full description on the dataset page: https://huggingface.co/datasets/notgoodkeeper/cnn-based-drowsiness-detection-data.

sourceHugging Facemitupdated 1mo agoView on Hugging Face

1likes752downloads

Dataset Card

CNN-Based Drowsiness Detection - Dataset

Preprocessed, auto-labeled face-crop images used to train the model in notgoodkeeper/cnn-based-drowsiness-detection.

Code: https://github.com/not-good-keeper/cnn-based-drowsiness-detection

Collection

Frames were captured from a webcam, then run through:

1.Haar Cascade face detection -> crop + pad + resize to 412x412
2.MediaPipe Selfie Segmentation -> background replaced with white
3.CLAHE contrast normalization -> grayscale

MediaPipe FaceMesh (468 landmarks) was used to compute EAR, MAR, and head-pose (roll/pitch) per frame, and a hand-tuned heuristic scorer turned those into a weak drowsiness_class / drowsiness_score label. See src/feature_extraction.py and src/collect_dataset.py in the code repo for the exact logic.

Splits

split	rows
train	1,661
val	356
test	359

Columns

column	description
`image`	relative path to the preprocessed face image (`images/...png`)
`drowsiness_class`	heuristic-labeled class, 0-4 (alert -> extremely drowsy)
`drowsiness_score`	heuristic score, normalized 0-1
`ear`	eye aspect ratio, normalized 0-1
`mar`	mouth aspect ratio, normalized 0-1
`head_roll`, `head_pitch`, `head_combined`	head-pose angles, normalized 0-1

Regression columns are min-max normalized to [0, 1]; see the code repo's src/realtime_inference.py for the original-scale ranges used to denormalize them.