The dataset viewer is not available for this dataset.
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
HM3D Subset (3DVLM)
A small, fast-to-download slice of HM3D scenes, rendered into a uniform
posed-RGB-D format for quick model test-runs. This is a subset of the full
set: 100 scenes (randomly sampled, seed 0) out of 779, with all episodes
kept for each selected scene. For the complete data, see the full (private)
3dvlm-hm3d.
Contents
100 scenes, one .tar each under hm3d/. Each scene tar holds its episodes
(short 5-frame clips, 1–5 per scene → ~5–25 frames per scene). Each episode
extracts to its own directory:
00002-FxCkHAfgh7A_ep00/
├── images/ # frame_000000.jpg … frame_000004.jpg (5 RGB frames, 512×512)
├── depth.npy # (5, 512, 512) float32
├── valid_mask.npy # (5, 512, 512) bool — True where depth is valid
├── extrinsics.npy # (5, 4, 4) float32 — world→camera (w2c)
├── intrinsics.npy # (5, 3, 3) float32 — pinhole K
└── meta.json # scene_id, frame_ids (parallel to array index), image_size
An episode is a self-contained 5-frame clip with frame-to-frame overlap; treat
each as one short posed-RGB-D sequence. The first axis of every array is the frame,
matching meta.json's frame_ids and the sorted images/ files.
Conventions
- Coordinate frame: OpenCV (x-right, y-down, z-forward).
extrinsicsis the world→camera (w2c) matrix; invert it for camera→world. - Depth: z-depth in metres (distance along the camera z-axis, not Euclidean
ray length), returned natively by the simulator. Use
valid_maskfor valid pixels. - Intrinsics: fixed —
fx=fy=256,cx=cy=255.5(90° FOV, 512×512).
Quick start
import tarfile, json, numpy as np
from huggingface_hub import hf_hub_download
p = hf_hub_download("helioom/3dvlm-hm3d_subset", "hm3d/00002-FxCkHAfgh7A.tar", repo_type="dataset")
tarfile.open(p).extractall("hm3d/")
ep = "hm3d/00002-FxCkHAfgh7A_ep00"
meta = json.load(open(f"{ep}/meta.json"))
depth = np.load(f"{ep}/depth.npy") # (5, 512, 512)
K = np.load(f"{ep}/intrinsics.npy") # (5, 3, 3)
w2c = np.load(f"{ep}/extrinsics.npy") # (5, 4, 4)
License
Built on HM3D (Matterport), released for research use only under the Matterport end-user license; the same terms apply to this derived subset. See HM3D.
- Downloads last month
- 16