| --- |
| license: other |
| task_categories: |
| - image-segmentation |
| tags: |
| - robotics |
| - video |
| - foreground-segmentation |
| --- |
| |
| # RoboMME demonstration-prefix frames with foreground masks |
|
|
| The **demonstration prefix** of RoboMME episodes -- the video-instruction frames, where `info/is_video_demo` |
| is True -- together with a binary foreground mask for each frame. |
|
|
| These frames are worth publishing separately because they are missing from the usual export: the pickle |
| conversion that training pipelines consume writes a file only when `is_demo` is False, so the demo prefix |
| (292k of RoboMME's 769k frames) exists only in the raw HDF5. |
|
|
| Of the 16 tasks, **9 carry a demo prefix and 7 do not**. Both are here, kept apart because the frames mean |
| different things: |
|
|
| | | tasks | episodes | frames | what the frames are | |
| |---|---|---|---|---| |
| | `<Task>/` | 9 | 2 each | 6,024 | the demo prefix only | |
| | `no_demo_tasks/<Task>/` | 7 | 2 each | 6,010 | whole episodes -- there is no demo portion to cut | |
|
|
| With a demo: InsertPeg, MoveCube, PatternLock, RouteStick, VideoPlaceButton, VideoPlaceOrder, VideoRepick, |
| VideoUnmask, VideoUnmaskSwap. Demo length splits by family -- the `Video*` tasks run long (VideoPlaceOrder |
| 940 and 1105 frames) while others are short (PatternLock 48 and 63). |
|
|
| Without: BinFill, ButtonUnmask, ButtonUnmaskSwap, PickHighlight, PickXtimes, StopCube, SwingXtimes. |
|
|
| ## Contents |
|
|
| ``` |
| <Task>/episode_<NNNN>/rgb/00000.png front camera, 256x256 RGB, raw pixels |
| <Task>/episode_<NNNN>/mask/00000.png 1-bit foreground mask, same size |
| <Task>/episode_<NNNN>/done.json kind, has_demo, demo_len, num_timesteps, |
| per-frame foreground fraction |
| no_demo_tasks/<Task>/episode_<NNNN>/ same layout, whole episodes |
| demo_manifest.json episode list for the demo tasks |
| no_demo_manifest.json episode list for the demo-less tasks |
| ``` |
|
|
| Frame `00000` is absolute timestep 0 of the episode and frames run to `demo_len - 1`, which is the end of the |
| demo prefix under `<Task>/` and the end of the episode under `no_demo_tasks/`. Check `kind` |
| (`demo` / `full_episode`) in `done.json` rather than inferring from the path. RGB is untouched: no crop, no |
| resize, no normalisation, so any preprocessing is the consumer's choice. |
|
|
| `is_video_demo` was verified to be a clean prefix -- True..True then False..False -- by reading every timestep |
| of every exported episode. |
|
|
| ## Masks |
|
|
| `ZhengPeng7/BiRefNet` (the full model), run at 1024x1024 and thresholded at 0.5, then resized back to 256. The |
| resolution matters: below roughly 512 the mask comes back essentially empty, which makes a |
| foreground-weighted loss put most of its weight on the wooden table instead of the arm and the objects. |
|
|
| The masks are model output, not ground truth. They are accurate on the arm and the manipulated objects |
| and they do pick up small painted markers, but they are not a substitute for simulator segmentation. |
|
|
|
|
| These masks were regenerated with the **full** BiRefNet, replacing an earlier set from |
| `BiRefNet_lite`. Same resolution, same 0.5 threshold, same resize back -- only the checkpoint |
| changed. The lite model drops the robot arm entirely on some frames: over the 12,034 frames |
| here the full model finds more foreground in 58.2% of them and changes the mask by more than |
| 2% of the frame in 5.5%, and the worst case (VideoPlaceButton episode_0000 frame 00304) goes |
| from a foreground fraction of 0.032 to 0.186 because lite masked the buttons and the cube but |
| not the arm holding them. Mean foreground fraction rose from 0.0795 to 0.0845; per task the |
| shift ranges from -0.0002 (PatternLock) to +0.0176 (BinFill). |
| |
| `done.json`'s `foreground_fraction`, `seg_model` and `seg_input_size` were recomputed to match. |
| |
| ## Provenance and licence |
| |
| Frames are derived from the RoboMME dataset; its licence governs their use. The masks are generated by |
| BiRefNet_lite (MIT). Nothing here is human-annotated. |
|
|
