Datasets:

G-G-Games
/

diecamera-crops

license: other
license_name: cc-by-nc-4.0-ours-cc-by-4.0-roboflow
license_link: https://github.com/eschatus/diecamera/blob/main/LICENSE-NC.md
task_categories:
  - image-classification
language:
  - en
tags:
  - dice
  - polyhedral-dice
  - ttrpg
  - tabletop
  - dice-reading
pretty_name: dieCamera dice crops
size_categories:
  - n<10K

dieCamera — per-die crops

One cropped image per physical die, labelled with its type and face value. This is the deliberately-simple training set for dieCamera's offline value reader — the app that watches a dice tray and posts the roll into a virtual tabletop (source).

For the full frames these crops were cut from (and the multi-die detector-training data), see the companion repo G-G-Games/diecamera-frames.

Schema

Standard 🤗 imagefolder layout — load_dataset("G-G-Games/diecamera-crops") needs no config.

data/<file>.jpg        one die, cropped to its bounding box + a small margin
data/metadata.jsonl    one row per crop

column	type	meaning
`file_name`	string	the crop image
`source`	string	`rig` (our webcam) or `roboflow:<fork>` (a third-party image, see below)
`type`	string	die type — `d4`, `d6`, `d8`, `d10`, `d12`, `d20`
`value`	int	the up-face value read (d10 may be 0)
`date`	string	capture date (`rig`), or publish date when the source has none
`added_at`	string	the day this crop first entered the dataset — filter `added_at > last_run` to train only on what's new
`holdout`	bool	`true` = reserved for evaluation; filter these out when training

Training tip. Exclude eval frames and (optionally) skip what you've already trained on: ds.filter(lambda r: not r["holdout"]). added_at lets an incremental finetune pick up only rows added since its last run, instead of reprocessing the whole set.

Every die here has a trusted face value: rig dice are human-confirmed or placed to a prompt; roboflow dice are the ones a human reviewed and confirmed by hand.

Licence — read this before commercial use

This dataset is mixed-licence, and the source column tells you which applies per row:

`source`	licence
`rig`	CC BY-NC 4.0 © G-G-Games — free personal use; commercial needs a licence
`roboflow:<fork>`	CC BY 4.0 © the fork's original author (commercial OK, attribution required)

The labels on every row are G-G-Games' own work (CC BY-NC 4.0). Full terms: LICENSE-NC.md.

Roboflow attribution (CC BY 4.0)

Crops with a roboflow: source derive from these Roboflow Universe datasets, used with modifications (cropped; our own top-face type/value labels added):

`source`	original author → dataset
`roboflow:200_dataset`	vkr-55xr7 → https://universe.roboflow.com/vkr-55xr7/200_dataset
`roboflow:d4-turbo-rad-v4`	turbo-rad → https://universe.roboflow.com/turbo-rad/d4-bmzdm

See ATTRIBUTIONS.md for the full provenance record.