PQS / README.md
perspectivequadrupletsuite's picture
Add files using upload-large-folder tool
ac6a695 verified
|
Raw History Blame Contribute Delete
6.08 kB
---
license: cc-by-4.0
pretty_name: Perspective Quadruplet Suite (PQS)
language:
- en
task_categories:
- visual-question-answering
tags:
- spatial-reasoning
- perspective-taking
- vision-language-models
- interpretability
- synthetic
size_categories:
- 10K<n<100K
configs:
- config_name: synthetic
data_files:
- split: all
path: synthetic/items.jsonl
- config_name: real_sunrgbd
data_files:
- split: all
path: real_sunrgbd/items.jsonl
- config_name: control
data_files:
- split: all
path: control/items.jsonl
---
# Perspective Quadruplet Suite (PQS)
Released anonymously for review (CVPR 2027 submission).
PQS tests whether a vision-language model can answer a spatial question from a viewpoint
other than the camera's. Every image is asked two questions that differ in one clause:
one from the camera's point of view and one from the point of view of a visible observer
(a humanoid, a chair or a tripod camera). Both questions point at the **same image file**.
When the two viewpoints give different answers the pair is a **conflict** pair, otherwise
an **agree** pair; on conflict pairs, an answer given from the camera's viewpoint is wrong.
## Contents
| folder | what | size |
|---|---|---|
| `synthetic/` | 8,000 quadruplets, 46,804 questions, 40,000 Blender/Cycles renders (448x448) | ~9 GB |
| `real_sunrgbd/` | 356 camera/observer question pairs on 180 SUN RGB-D photographs (seats as observers), 2,136 question rows including blind and text-only controls | <3 MB |
| `control/` | 1,351 scene pairs (2,702 questions) where the camera is re-placed at the observer's pose, so the observer question becomes an ordinary camera question with the same answer | <7 MB |
`SHA256SUMS` lists the hash of every file.
### synthetic/
- `items.jsonl`: one row per question.
- `scenes.jsonl`: one row per render, with camera pose, observer pose and heading, object positions and room bounds.
- `render_meta.jsonl`: one row per render, with the image file name, its SHA-256 and the 2D boxes of the questioned objects. The `image_path` prefix is the directory used at render time; only the file name matters.
- `meta.json`: generation settings (seed 20260829) and the planned cell counts.
- `splits.csv`: the train / dev / test split used for the intervention experiments (3,960 / 791 / 3,249 quadruplets), assigned by scene so that no layout appears in two splits. It is `u = int(sha256("EXP4-A9|" + quadruplet_id)[:12], 16) / 16**12`, with train `u < 0.5`, dev `0.5 <= u < 0.6` and test otherwise.
- `renders/renders_full_part{0..3}.tgz`: the 40,000 PNGs, 10,000 per archive, named `<quadruplet_id>_<role>.png`.
Render roles, 8,000 each:
- `base`: the camera and observer questions.
- `rotated`: observer heading turned 90 degrees.
- `mirrored`: observer position reflected across the camera axis.
- `observer_absent`: the observer is not drawn, and its pose is given in words.
- `calibration`: a reference render of the same layout.
Each quadruplet also has text-only and blind variants with probability 0.25.
### real_sunrgbd/
The photographs are **not** included. Obtain SUN RGB-D from its original source and under its own terms. Each row gives `source_sequence`, the path inside the SUN RGB-D release, and `image_sha256`, so you can check that you have the same image. Geometry comes from the SUN RGB-D 3D box annotations: the camera is the origin, a seat is the observer, and the seat's front is its heading.
### control/
These files hold questions and scene geometry only; the renders are not included in this release. Both questioned objects lie inside both views' fields of view, with a 3-degree margin. Disparities are 90, 135 and 180 degrees.
## Question fields
| field | meaning |
|---|---|
| `quadruplet_id`, `scene_id` | quadruplet and render identifiers |
| `relation` | `left_right`, `front_behind`, `closer_farther` or `facing_toward` |
| `frame` | `camera` or `observer` |
| `condition` | `conflict`, or `coincide` (agree) |
| `role`, `variant` | render role; `normal`, `text_only` or `blind` |
| `prop` | observer prop: `humanoid`, `chair` or `tripod_camera` |
| `theta_deg` | camera-observer heading disparity: 0, 45, 90, 135 or 180 |
| `question`, `options` | question text and the two answer options |
| `answer_idx`, `answer_label` | correct option for the asked frame |
| `egocentric_answer_idx`, `egocentric_label` | the answer from the camera's viewpoint |
| `margin` | geometric margin of the label; near-ties were removed |
`left_right` is defined by bearing (the angle at which an object appears from the viewpoint), which agrees exactly with the rendered projection.
> **Important: `front_behind` labels are stored inverted.** The generator calls the subject "front" when it lies farther along the viewer's heading than the reference object, which in English means *behind*. Every result in the paper uses the corrected label. For `front_behind` rows, in all three folders, the correct answer to the English question is the **other** option:
>
> ```python
> correct_idx = 1 - row["answer_idx"] if row["relation"] == "front_behind" else row["answer_idx"]
> camera_idx = 1 - row["egocentric_answer_idx"] if row["relation"] == "front_behind" else row["egocentric_answer_idx"]
> ```
>
> `condition` (conflict or agree) is unaffected, because it compares two labels that are both inverted. The files are left unchanged so that their hashes match those reported in the paper.
## Scoring used in the paper
Each question is scored by comparing the logits of the two option letters, with four phrasings and both option orders. Three accuracies are reported:
- `M`: accuracy on camera questions.
- `C`: accuracy on observer questions in conflict pairs.
- `K`: accuracy on observer questions in agree pairs.
The summary score is balanced accuracy `(C + K) / 2`. Always answering from the camera gives C = 0 and K = 1, and so scores 0.5.
## License
The synthetic data, all annotations and all question files are released under CC BY 4.0. SUN RGB-D images are not redistributed here and remain under their original terms.