Stefanobraghetto/dreamer4-masked-roi-180

Dreamer4 Masked ROI Dataset (all-37, 180px)

Per-frame frames + binary ROI masks for DeepMind-Control-style episodes, used to train a foveated (coarse/fine) video tokenizer for a Dreamer4-style world model. The ROI mask marks the agent / region of interest in each frame.

Provenance & license (please read)

This dataset is derived, and the uploader does not own or claim copyright over the underlying frames. The RGB frames come from rollouts generated by a third-party, unofficial Dreamer4 implementation of DeepMind Control Suite (dm_control / MuJoCo) environments; only the binary ROI masks were produced by the uploader. The license is therefore marked unknown: no warranty is given, and no MIT or other permissive grant is implied. Before redistributing or using this data (especially commercially), verify the licenses of the upstream sources:

• DeepMind Control Suite / dm_control and MuJoCo (the environments rendered).
• The specific unofficial Dreamer4 implementation used to generate the rollouts.

If you are a rights holder and want attribution corrected or the data removed, please open a discussion on this repo. This repository is shared for research reproducibility only.

Format

WebDataset tar shards, one tar per episode, laid out as <level>/<episode>.tar with an <episode>.metadata.json sidecar. Each tar contains, per frame:

• NNNNNN.frame.png — RGB frame, 180x180
• NNNNNN.mask.png — binary ROI mask, 180x180 (0 / 255)

Levels: expert, mixed-small, mixed-large. ~972 episodes, ~3.75M frames. overlay_frames, masked_frames and strip images are not included (they are derivable from frames + masks).

Loading

from datasets import load_dataset  # pip install datasets webdataset
ds = load_dataset(
    "webdataset",
    data_files={"train": "hf://datasets/Stefanobraghetto/dreamer4-masked-roi-180/**/*.tar"},
    streaming=True,
)
for s in ds["train"]:
    frame, mask = s["frame.png"], s["mask.png"]  # PIL images
    break