kieranpattison/visual-perspective-taking

Occlusion Deception Visual Perspective-Taking Pilot

This synthetic image dataset asks whether a visible robot can see a named target object from a third-party viewpoint. Each example has one RGB image, a question, three answer options, and one of three labels: yes, no, or cannot_tell.

Dataset structure

The single train split contains 100 examples generated from ProcTHOR houses 64, 461, 401, 241, 634 with seed 43. Label counts are: yes=59, no=1, and cannot_tell=40.

The columns are:

• image: decoded automatically from file_name by Hugging Face ImageFolder.
• id: deterministic example identifier.
• question, option_a, option_b, option_c: the VQA prompt and choices.
• answer, ground_truth_choice, label, label_id: answer representations. Label IDs are yes=0, no=1, and cannot_tell=2.
• target_type, house_index, room_id: limited scene descriptors.
• policy_version, sampling_policy_version: labeling and sampling provenance.

Upload

Create an empty dataset repository on the Hugging Face Hub, authenticate locally, then run this from the directory containing this README:

python -m pip install -U huggingface_hub
hf auth login
hf upload YOUR_USERNAME/YOUR_DATASET . . --repo-type dataset

Load a local copy with:

from datasets import load_dataset

dataset = load_dataset("imagefolder", data_dir=".")

After upload, replace the path with the Hub repository ID:

dataset = load_dataset("YOUR_USERNAME/YOUR_DATASET")

Label semantics

• yes: the available image evidence is sufficient to establish that A can see B.
• no: the available image evidence is sufficient to establish that A's sight bundle is blocked.
• cannot_tell: the image does not expose enough evidence to establish either conclusion.

The label policy is vpt-label-v2.5-epistemic-rays. Labels are generated from simulator geometry and C-visible depth evidence, not from human annotation.

Scope and limitations

This is a 100-example pilot from five synthetic houses, not an unconditional sample of indoor scenes. Its label distribution is induced by the repository's documented target-facing sampling policy. It should not be treated as a representative real-world benchmark without further study.

Only public RGB images and task metadata are included. Private depth images, instance masks, agent poses, ray evidence, world-truth diagnostics, and audit overlays are intentionally excluded.

The dataset author has not yet specified a distribution license. Review the AI2-THOR and ProcTHOR terms and select an appropriate dataset license before making the Hub repository public.