yinbq/sceneshift-InteriorGS-50

---
license: mit
task_categories:
- visual-question-answering
- image-text-to-text
language:
- en
tags:
- 3d-scene
- spatial-reasoning
- camera-pose
- benchmark
- gaussian-splatting
size_categories:
- 100K<n<1M
---

# SceneShift InteriorGS Bench (50 Scenes)

A spatial reasoning VQA benchmark built on 3D Gaussian Splatting (3DGS) reconstructions of real indoor scenes from the InteriorGS dataset.

## Dataset Summary

- **50 indoor scenes** from InteriorGS
- **129,362 questions** across 13 question types
- **3,820 rendered images** from diverse camera poses
- **4 camera motion patterns**: around, linear, rotation, spherical
- **13 question types**: object_size, object_distance_to_camera, object_pair_distance_center, relative_size, relative_distance, relative_distance_to_camera, mc, object_size_comparison_relative, object_size_comparison_absolute, object_pair_distance_center_w_size, object_pair_distance_vector, object_comparison_absolute_distance, object_comparison_relative_distance

## Directory Structure

```
{pattern}/{scene_id}/
├── metadata.json # Scene metadata (objects, camera count, question types)
├── questions.jsonl # QA pairs (one JSON object per line)
└── images/
├── pose_0000.png # Rendered views from different camera poses
├── pose_0001.png
└── ...
```

**Patterns**: `around/`, `linear/`, `rotation/`, `spherical/`

## Question Format (questions.jsonl)

Each line is a JSON object:

```json
{
"question": "What is the estimated length of the chair in this scene in meters? ...",
"answer": "1.26",
"answer_type": "numerical",
"question_type": "object_size",
"question_id": "object_size_29_length",
"primary_object": "29",
"objects": [...],
"camera_pose_index": 0,
"image_path": "pose_0000.png"
}
```

## Question Types

| Type | Description |
|------|-------------|
| `object_size` | Estimate object dimensions (length/width/height) |
| `object_distance_to_camera` | Distance from an object to the camera |
| `object_pair_distance_center` | Distance between two object centers |
| `relative_size` | Yes/No comparison of two objects' sizes |
| `relative_distance` | Yes/No comparison of two distances |
| `relative_distance_to_camera` | Yes/No: which object is closer to camera |
| `mc` | Multiple-choice (A/B/C/D) spatial questions |
| `object_size_comparison_relative` | Relative size ratio between objects |
| `object_size_comparison_absolute` | Absolute size difference between objects |
| `object_pair_distance_center_w_size` | Distance considering object sizes |
| `object_pair_distance_vector` | Directional distance vector between objects |
| `object_comparison_absolute_distance` | Absolute distance difference |
| `object_comparison_relative_distance` | Relative distance ratio |

## Citation

If you use this dataset, please cite the associated work.