HM3D-M3 data v1.0.0

HM3D-M3 data release v1.0.0

Assets for the HM3D-M3 benchmark. Code: https://github.com/NJU-R-L-Group-Embodied-Lab/hm3d-m3-code. Train split and the 100K
teacher samples: https://huggingface.co/datasets/NJU-R-L-Group-Embodied-Lab/HM3D-M3.

asset	content
ma_official_val_1000_samefloor_mixed3_seed1123.json.gz	Val-1000 episodes (1,000 episodes, 3,926 goals, 36 HM3D val scenes)
ma_official_mini_100_samefloor_mixed3_kmquota_seed1123.json.gz	Mini-100 episodes: the original 100 Mini-100 ids, copied from Val-1000 (strict subset)
iin_images_official_val1000_mixed3_stretch512.tar.gz	1,314 IIN goal images, stretch512 (default; LaViRA rows)
iin_images_official_val1000_mixed3_stretch_v2.tar.gz	1,314 IIN goal images, stretch_v2 (the four adapted OD baselines on Val-1000)
iin_images_official_mini100_mixed3_stretch512.tar.gz	133 IIN goal images for Mini-100 (subset of the Val stretch512 pack)
RELEASE_NOTES.md	these notes
SHA256SUMS	checksums of the assets above

Install

From the sim/ directory of the code repository:

D=data/datasets/ma_objectnav_ovon/val_unseen && mkdir -p $D
sha256sum -c SHA256SUMS
gunzip -c ma_official_val_1000_samefloor_mixed3_seed1123.json.gz > $D/ma_official_val_1000_samefloor_mixed3_seed1123.json
gunzip -c ma_official_mini_100_samefloor_mixed3_kmquota_seed1123.json.gz > $D/ma_official_mini_100_samefloor_mixed3_kmquota_seed1123.json
for f in iin_images_*.tar.gz; do tar -xzf "$f" -C $D; done

Each image archive extracts to one folder with the PNGs and one manifest.json
(image paths are relative to sim/; file is relative to the manifest; each item
has sha256 and the render camera). Byte-identical images are stored once and
extracted as hard links.

Changes with respect to the files used during development

• Mini-100 is re-exported from Val-1000, so it is a strict subset. The development
  Mini-100 file had independently regenerated TN descriptions: 108 of 127 tn_text
  values differ in wording; all other fields were already identical.
• Every IIN goal has a new iin_render_camera field (hfov, width, height,
  camera_height_m, standoff_m, variant) with the camera of the stretch512 images.
  iin_viewpoint is kept unchanged; it is source IIN metadata, not the render camera.
• One merged manifest.json per image folder, relative paths only. In the stretch_v2
  manifest, 133 items without camera metadata were completed from the identical
  Mini-100 stretch_v2 renders.

License

CC BY-NC 4.0 for annotations; HM3D scenes are required and the images are HM3D
renderings subject to the Matterport academic EULA. See Part 2 of LICENSE in the code
repository.