Datasets:
Access LOOP-Bench
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
LOOP-Bench contains RGB-D images rendered from HM3D-Semantics scenes (Matterport). Before downloading, you must have accepted the Matterport terms of use for HM3D (https://matterport.com/habitat-matterport-3d-research-dataset), and you may use this dataset only for non-commercial research under those terms.
Log in or Sign Up to review the conditions and access this dataset content.
LOOP-Bench: Long-Term Object Occurrence Prediction
LOOP-Bench tests whether a robot can learn where an object usually occurs in a home from many traversals in which it sees that object only sometimes. Over each traversal, objects are placed again at random from hidden long-term distributions. The robot's routes and viewpoints are sampled independently of those placements, so how often an object occurs and how often the robot could observe it vary separately. Queries are given only after all the traversals are recorded. The benchmark measures two things: how well a method recovers each object's occurrence distribution, and how well that estimate guides an active search in the simulator.
It is the benchmark of Retrospective Open-Vocabulary Memory for Long-Term Object Search (Wang, Xue, Chen, Peng,
Soh; National University of Singapore). The code (ECROM, the baselines and the evaluation) is at
https://github.com/jiaming-ai/ecrom. Its scripts/fetch_loopbench.sh downloads this dataset and puts every file
where the code expects it.
Contents
| Homes | 10 HM3D-Semantics v0.2 houses (8 train, 2 val split), 16–33 annotated receptacle spots each |
| Objects | 8 Google Scanned Objects per house, each with a hidden placement distribution over the spots (unimodal, bimodal, near-uniform or rare/absent) |
| Mapping history | 20 continuous traversals per house (200 in total, 79,741 RGB-D frames) of varying coverage, from broad routes to short local patrols |
| Queries | 3 calibration objects and 5 held-out query objects per house (50 queries in total) |
| Search | 150 test cases (5 queries × 3 placements × 10 houses, 128 with the object present), 120 m budget |
| Sensor | 800×600 RGB (JPEG) and depth (16-bit PNG, millimetres), 60° horizontal FoV, camera 1.0 m above the floor, a world-frame camera pose for every frame |
Layout
houses/hm_<id>.tar one house; extract into data/hm3d_v10_20/
hm_<id>/
scene.json HM3D scene id, camera intrinsics, stations, receptacle spots, objects with their
hidden placement distributions (gt.pi over spots), object scale
DATASET_MANIFEST.json the 20 mapping traversals, frame counts and episode checksums
epNNN/episode.json one traversal: route, object placement and poses, per-frame camera pose T_wc,
discrete actions. Search-test episodes have no frames.
epNNN/rgb/00000.jpg RGB frames (mapping traversals only)
epNNN/depth/00000.png depth in millimetres (uint16)
map/ the CROSS keyframe map built from the 20 traversals (map.pkl), the place-view atlas,
the surface-patch supports (atlas_patch_odom.pkl) and the stored keyframes' OWLv2 scores
benchmark/final/ the frozen benchmark definition; place it in aggregate_out_v10/pooled_20260925/final/
design.json selected traversals per house, the 150 search cases and the evaluation settings
priors/hm_<id>.npz every method's frozen prediction per query and the support-level ground truth
prepared/hm_<id>.json per-house query and calibration metadata
dataset_manifests/ a copy of each house's DATASET_MANIFEST.json
summary.json the reference results: aggregate metrics and per-house numbers
development_trials.json the reference results: outcomes of all 14 methods on the 150 search cases
assets/gso_loopbench.tar the 27 Google Scanned Objects used by the scenes, converted for Habitat
(.glb, .object_config.json, .png); point ECROM_GSO to the extracted folder
SHA256SUMS checksums of every file above
map/*.pkl files are Python pickles: load them only from this repository.
Download
Click Agree and access on this page, log in with hf auth login, then from a clone of the code:
bash scripts/fetch_loopbench.sh # all ten houses (~15 GB)
HOUSES="hm_2Pc8 hm_DqJK" bash scripts/fetch_loopbench.sh # only some houses
Without the code: hf download Jamie-nus/LOOP-Bench --repo-type dataset --local-dir LOOP-Bench, then extract each
tar.
The search simulator also needs the HM3D-Semantics v0.2 scenes. Download them yourself under your Matterport
account (python -m habitat_sim.utils.datasets_download --username <api-token-id> --password <api-token-secret> --uids hm3d_train_v0.2 hm3d_val_v0.2) and set ECROM_HM3D to the folder that contains
hm3d_annotated_basis.scene_dataset_config.json.
License
The RGB-D frames and maps are rendered from HM3D scenes and fall under the Matterport terms of use for HM3D: non-commercial research only, and they may not be redistributed. The objects are from Google Scanned Objects (CC BY 4.0). The annotations, placements and benchmark definition are released for non-commercial research use.
- Downloads last month
- 39