Schrodin-purrrrr/toothbrush-v2-dataset

Toothbrushing Detection Dataset (v2)

Video and image data for detecting toothbrushing behavior, collected for a Raspberry Pi Zero 2W toothbrush-detection project (toothbrush_v2). A single-class object detector is trained on this data to output [x, y, w, h, confidence] for the toothbrush in frame.

Dataset structure

Files are packed into tar shards (rather than uploaded individually) to stay within the Hub's per-repo file-count guidelines. To reconstruct the original layout, extract every shard_*.tar in a folder into that folder, then use the matching manifest to sort files by split.

Path	Contents	Size
videos/*.mp4	63 source recordings, positive*.mp4 (toothbrushing) / negative*.mp4 (not toothbrushing)	1.8G
frames/shard_0000.tar … shard_0013.tar	4,253 frames extracted from the videos (positive/, negative/ subfolders)	2.8G
labels_frames/shard_0000.tar	YOLO-format labels for frames/	8.4M
aug_224_gray/shard_0000.tar, shard_0001.tar	25,266 augmented, resized-to-224, grayscale image+label pairs (images/, labels/)	478M
manifests/frames_split.json	Train/val split membership for each file in frames/	—
manifests/aug224_split.json	Train/val/test split membership for each file in aug_224_gray/	—

Label format

YOLO format: class x_center y_center width height, all normalized to [0, 1]. Single class (0 = toothbrush). Negative (no-toothbrush) frames use the point form 0 0.5 0.5 0.0 0.0.

Class balance

• frames/: 1,995 positive, 2,258 negative
• aug_224_gray/ split sizes: 17,682 train / 3,786 val / 3,798 test

Known issue: train/val overlap in frames_split.json

1,150 of the 4,253 frames appear in both the train and val portions of the original (pre-augmentation) split — i.e. there is genuine leakage in that split, preserved as-is in the manifest for transparency rather than silently fixed. manifests/frames_split.json stores a list of splits per file, so entries with ["train", "val"] are the leaked ones. The downstream aug_224_gray train/val/test split does not have this issue (verified no overlap).

If you're training/evaluating directly on frames/, either drop the leaked files from val or re-split before reporting held-out metrics.

Redundant folders that were dropped during packing

The original working directory also contained consolidated/, frames_train/, frames_val/, dataset_224_gray/, and split_224/. These were verified byte-for-byte duplicates of frames/ + labels_frames/ (for consolidated/) or aug_224_gray/ (for dataset_224_gray/), just reorganized into split subfolders — so they were left out of this upload to avoid uploading the same image content multiple times. Their split membership is fully captured in the two manifest files above.

License

Not yet specified. Update this field (and the YAML license: key above) before relying on this dataset outside the originating project — this data includes video of a person's face/mouth while brushing, so confirm any consent/PII handling requirements are met before broader release.

Source

Collected as part of a Raspberry Pi Zero 2W toothbrush-detection and brushing-quality-monitoring project. toothbrush_v2 is a historical generation of the system, superseded by later versions but kept for reference.