HUDseg

Lossless per-pixel semantic segmentation videos of VoxeLibre (Minetest) gameplay, where the segmentation also covers the first-person overlays: the HUD hotbar (hud) and the block held in hand (wielded_item). Generated from the seg10k dataset with tools/seg_add_overlays.py in dynamic-agent-dataset-generation.

Layout

<split>/<seed>/<seed>.mp4          # lossless segmentation video, one colour per class
<split>/<seed>/seg_classes.json    # class list for this level

split	levels
train	9,600
eval1	200
eval2	200

Classes and colours

Each seg_classes.json lists the classes present in that level, with name, kind (sky, node, overlay, ...), color (RGB), pixel_fraction and frames (number of frames the class appears in). Colours are deterministic across levels:

• sky = [135, 206, 235]
• any other class = first 3 bytes of md5(f"{name}#{salt}"), using the smallest salt >= 0 whose colour is at least 48 (RGB distance) away from the sky colour.

Videos are lossless, so a pixel's class is recovered exactly by matching its RGB value against the level's class list.

License

MIT