Datasets:
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains information derived from the Habitat-Matterport 3D Research Dataset (HM3D) and is available only for non-commercial academic use. Access is governed by the Matterport End User License Agreement for Academic Use of Model Data. By requesting access, you affirm that you have read and accept that agreement.
Log in or Sign Up to review the conditions and access this dataset content.
Occlusion Deception HM3D Visual Perspective-Taking
This ImageFolder dataset asks whether a visible robot can see a red target box from a
third-party camera. Each example contains one RGB image, a question, three answer options,
and an evidence-based yes, no, or cannot_tell label.
Access and permitted use
The images are Matterport Dataset Derived Information. Access and use are limited to non-commercial academic purposes under the Matterport End User License Agreement for Academic Use of Model Data. Do not attempt to identify the address, owner, occupants, or location represented by any environment. A is our original box-built quadruped proxy, with independently posed head and torso. It is not a manufacturer robot model. The target model comes from the YCB Object and Model Set, licensed CC BY 4.0.
This dataset repository uses manual gating. Requesters must provide their academic institution, state their intended non-commercial academic use, and affirm acceptance of the Matterport agreement. Those restrictions apply to downloaded copies and derived information.
Contents
The train split contains 1000 generated examples across 10 HM3D v0.2 minival
environments. The split name is an ImageFolder packaging convention. Labels were naturally
sampled rather than balanced: yes=544,
no=239, and
cannot_tell=217.
Only C's public 1024x1024 RGB frame and task metadata are distributed. Private A-view images, depth, semantic frames, exact poses, scene identifiers, navigation maps, heading sheets, visibility rays, and generation diagnostics are excluded.
Sampling and labels
C is placed on navigable floor. For A and separately for B, test up to five random horizontal rays with independent 5–25 m radii and retain the farthest actual traveled distance. A full-length ray ends its group early. B candidates remain at least 10 degrees from the selected A direction. A and B are placed within 0.5 m of their respective endpoints, with no A–B reachability constraint. A uses a 90-degree square field of view and faces B with small yaw jitter. Its head tilts vertically toward B within its joint limits. C must visibly identify both endpoints before the sample is retained.
The labeling policy is vpt-hm3d-v13-resolution-independent-visibility. It evaluates a 400-ray A-to-B sight bundle using simulator
geometry plus evidence visible from C. Missing depth is ignored; unresolved geometry and
out-of-frame evidence remain unknown.
The answer requires at least 10% of the A-to-B sight bundle to be clear. The box's pixel size in A's private render does not determine the public answer or force uncertainty. Private A-view renders are used only to validate definite answers against the same visible-fraction criterion.
Columns
image: C's public RGB image, decoded fromfile_nameby ImageFolder.id: deterministic identifier derived from the public image checksum.question,option_a,option_b,option_c: the task prompt and choices.answer,ground_truth_choice,label,label_id: answer representations, with label IDsyes=0,no=1, andcannot_tell=2.environment_index: a batch-local opaque environment index; source HM3D IDs are withheld.policy_version,sampling_method: generation provenance.
Citation
If you use these images, comply with the Matterport agreement and cite HM3D:
@inproceedings{ramakrishnan2021hm3d,
title={Habitat-Matterport 3D Dataset (HM3D): 1000 Large-scale 3D Environments for Embodied AI},
author={Ramakrishnan, Santhosh Kumar and others},
booktitle={NeurIPS Datasets and Benchmarks},
year={2021}
}
The YCB target is based on the object set introduced by Calli et al., "The YCB object and Model set: Towards common benchmarks for manipulation research," ICAR 2015.
- Downloads last month
- 571