README.md

---
license: cc-by-nc-4.0
pretty_name: SpatialMosaic VQA
task_categories:
- visual-question-answering
language:
- en
tags:
- spatial-reasoning
- vision-language
- vqa
- multi-view
- video-vqa
- partial-visibility
- occlusion
- low-overlap
- scannetpp
- waymo
- object-counting
- object-localization
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: indoor_test
    path: merged_indoor_test.json
  - split: outdoor_test
    path: merged_outdoor_test.json
---

<p align="center">
  <a href="https://arxiv.org/abs/2512.23365"><img src="https://img.shields.io/badge/arXiv-2512.23365-b31b1b.svg" alt="arXiv"></a>
  <a href="https://kanghee-lee.github.io/spatialmosaic/"><img src="https://img.shields.io/badge/Project-Page-blue" alt="Project Page"></a>
  <a href="https://github.com/Kanghee-Lee/SpatialMosaic-project"><img src="https://img.shields.io/badge/GitHub-SpatialMosaic--project-black?logo=github&logoColor=white" alt="GitHub"></a>
</p>

# SpatialMosaic: A Multi-View VLM Dataset for Partial Visibility

## Description

**SpatialMosaic** is a multi-view visual question answering dataset for evaluating spatial reasoning under **partial visibility**, **occlusion**, and **low-overlap views**. It pairs indoor ScanNet++ and outdoor Waymo scene references with multi-frame VQA annotations. Questions require models to combine fragmented evidence across 2-5 views, rather than answering from a single image. The tasks cover object counting, object presence, localization, best-view selection, object-object spatial relations, and view-specific position reasoning.

This Hugging Face repository contains **annotation files only**. It does not redistribute ScanNet++, Waymo, image, video, depth, or scene data. Users must obtain any required source images from the official dataset providers and comply with their licenses, access requirements, and usage terms.

The current configuration contains **2,200 test examples**: 1,100 indoor ScanNet++ examples and 1,100 outdoor Waymo examples. Across the two JSON files, the annotations include 17 `question_type` labels, 2-5 referenced frames per example, 169 indoor ScanNet++ scenes, and 9 outdoor Waymo scenes. Evaluation is performed by exact multiple-choice answer matching against `mc_answer`.

## Available Files and Splits

This card is configured for two annotation-only test splits: one indoor split based on ScanNet++ scene references and one outdoor split based on Waymo scene references.

| Split | Source reference | File | Examples / entries | Status |
| --- | --- | --- | ---: | --- |
| `indoor_test` | ScanNet++ | `merged_indoor_test.json` | 1,100 | Current config |
| `outdoor_test` | Waymo | `merged_outdoor_test.json` | 1,100 | Current config |

## Loading

```python
from datasets import load_dataset

ds = load_dataset("jmkey/spatial_mosaic_vqa", "default")
print(ds)
print(ds["indoor_test"][0])
print(ds["outdoor_test"][0])
```

For direct local use of the raw JSON files:

```python
import json

with open("merged_indoor_test.json", "r") as f:
    indoor_rows = json.load(f)

with open("merged_outdoor_test.json", "r") as f:
    outdoor_rows = json.load(f)
```

## Data Format

Each entry contains the referenced source dataset, scene name, frame identifiers, question type, question text, answer options, and the correct multiple-choice answer in `mc_answer`.

Common fields include:

- `dataset`: source dataset identifier, such as `scannetpp` or `waymo`
- `scene_name`: source scene or segment identifier
- `frames`: list of 2-5 frame identifiers used by the question
- `question_type`: task subtype label
- `question`: natural-language question
- `options`: multiple-choice answer candidates
- `mc_answer`: correct option label

Some entries include additional metadata such as `ground_truth`, `bbox_2d`, `bbox_2d_diag`, `GT Scenario`, `overlap_avg`, `occlusion_avg`, `occ_level`, `overlap_level`, and `vis_level`.

## Intended Use

SpatialMosaic is intended for non-commercial research on:

- multi-view visual question answering
- spatial reasoning under partial visibility
- occlusion-aware and low-overlap view understanding
- VLM evaluation and instruction-tuning on multi-frame inputs

Out-of-scope uses include commercial use under this annotation license, redistribution of restricted source datasets, and treating these annotations as a substitute for the original ScanNet++ or Waymo data.

## Evaluation

SpatialMosaic uses multiple-choice evaluation. The primary metric is exact-match accuracy against `mc_answer`.

```python
accuracy = sum(pred == row["mc_answer"] for pred, row in zip(predictions, rows)) / len(rows)
```

## Citation

If you use SpatialMosaic, cite the paper:

```bibtex
@article{lee2025spatialmosaic,
  title={SpatialMosaic: A Multiview VLM Dataset for Partial Visibility},
  author={Lee, Kanghee and Lee, Injae and Kwak, Minseok and Hong, Jungi and Ryu, Kwonyoung and Park, Jaesik},
  journal={arXiv preprint arXiv:2512.23365},
  year={2025}
}
```

## License

The released VQA annotations are licensed under **CC BY-NC 4.0**. This license applies only to the annotations. Source images, videos, depth, and scene data remain governed by the original ScanNet++ and Waymo licenses and terms.