README.md

---
pretty_name: VSI-Super-Wild
language:
- en
task_categories:
- visual-question-answering
- multiple-choice
tags:
- spatial-supersensing
- video-understanding
- long-video
- spatial-reasoning
- benchmark
size_categories:
- "1K<n<10K"
---

# [ECCV 2026] Towards Spatial Supersensing in the Wild

Tsinghua University, NVIDIA, Stanford University

[![Project Page](https://img.shields.io/badge/Project-Page-4B8BBE?logo=googlechrome&logoColor=white)](https://vsi-super-wild.github.io/)
[![Paper](https://img.shields.io/badge/Paper-arXiv-B31B1B?logo=arxiv&logoColor=white)](http://arxiv.org/abs/2607.13681)
[![Code](https://img.shields.io/badge/Code-GitHub-181717?logo=github&logoColor=white)](https://github.com/THUSI-Lab/VSI-Super-Wild)
[![Dataset](https://img.shields.io/badge/Dataset-Hugging%20Face-FFD21E?logo=huggingface&logoColor=black)](https://huggingface.co/datasets/THUSI-Lab/VSI-Super-Wild)

Humans make sense of continuous sensory streams by maintaining implicit world states that support spatial reasoning and prediction. **VSI-Super-Wild** studies whether multimodal models can develop similar spatial supersensing capabilities in genuinely long-form, **in-the-wild** videos.

The benchmark moves beyond short indoor clips and object-centric settings. It evaluates world modeling over the **agent-object-environment** triad with human-verified question-answer pairs grounded in real-world video streams.

![VSI-Super-Wild teaser](assets/teaser.png)

## Highlights

- **In-the-wild long-video benchmark.** VSI-Super-Wild contains 442 view-level videos across 8 scene categories, totaling 284.52 hours, with videos up to 261.08 minutes long.
- **Human-verified spatial QA.** The release includes 6,980 question-answer pairs covering motion, place ordering, object ordering, and continuous object counting.
- **Multi-anchor task suite.** Four tasks probe agent, environment, and object states over long temporal horizons.
- **Ready for evaluation.** The video archives and metadata are released on this Hugging Face dataset page, while evaluation code is maintained in the [GitHub repository](https://github.com/THUSI-Lab/VSI-Super-Wild).

## Motivation

Existing spatial supersensing benchmarks are an important step toward testing implicit world modeling, but they often remain synthetic, household-centered, or object-centric. VSI-Super-Wild targets natural long-video streams and broader world-state coverage across agent, object, and environment anchors.

![Motivation](assets/motivation.png)

## Task Suite

VSI-Super-Wild evaluates four task families over the agent-object-environment triad:

| Task | Name | Anchor | Answer Format | Q&A Count |
|---|---|---|---|---:|
| `VMR` | Motion Orientation Recall | Agent | Multiple choice | 1,215 |
| `VPO` | Place Temporal Ordering | Environment | Multiple choice | 1,302 |
| `VOO` | Object Temporal Ordering | Object | Multiple choice | 3,350 |
| `VOC` | Continuous Object Counting | Object | Integer | 1,113 |

`VOO` contains two balanced subtypes: 1,675 first-appearance recall questions and 1,675 last-appearance recall questions.

![Task suite](assets/task_illustration.png)

## Dataset Construction

The benchmark is built through a semi-automatic pipeline with human-in-the-loop verification. We collect and filter in-the-wild panoramic videos, project panoramas into perspective views, derive temporal and spatial metadata, synthesize rule-based QA, and verify the resulting samples with rollback when metadata or QA needs refinement.

![Data construction](assets/data_construction.png)

## Benchmark Statistics

The released metadata file contains 6,980 human-verified Q&A rows associated with 442 view-level videos.

![Benchmark statistics](assets/benchstatistics.png)

## Files on Hugging Face

The Hugging Face repository contains the QA metadata file and video archives grouped by video duration bucket:

```text
VSI-Super-Wild/
|-- qa.jsonl
|-- videos_0_10_part_000.zip
|-- videos_10_30_part_000.zip
|-- videos_10_30_part_001.zip
|-- videos_30_60_part_000.zip
|-- ...
|-- videos_60_120_part_001.zip
|-- videos_120_plus_part_000.zip
|-- ...
`-- videos_120_plus_part_008.zip
```

| Duration Bucket | Archive Pattern | Files | Approx. Size |
|---|---|---:|---:|
| 0-10 min | `videos_0_10_part_*.zip` | 1 | 49.53 GB |
| 10-30 min | `videos_10_30_part_*.zip` | 2 | 267.49 GB |
| 30-60 min | `videos_30_60_part_*.zip` | 6 | 833.35 GB |
| 60-120 min | `videos_60_120_part_*.zip` | 2 | 265.15 GB |
| 120+ min | `videos_120_plus_part_*.zip` | 9 | 1249.97 GB |
| **Total video archives** | - | **20** | **2665.49 GB** |

The `video_name` field in `qa.jsonl` refers to the corresponding video file used by each question.

This dataset page hosts the released videos and QA metadata. For evaluation, please use the official [GitHub codebase](https://github.com/THUSI-Lab/VSI-Super-Wild), which maintains the task definitions, video resolver, metrics, and evaluator scripts.

## Citation

```bibtex
@inproceedings{VSI_Super_Wild,
  title={Towards Spatial Supersensing in the Wild},
  author={Tianjun Gu and Tianyu Xin and Kuan Zhang and Bowen Yang and Kok-Chung Chua and Peize Li and Xinran Zhang and Yupeng Chen and Qiyue Zhao and Qinlei Xie and Jianhang Liu and Yucheng Lu and Yinan Han and Marco Pavone and Yiming Li},
  booktitle={The Nineteenth European Conference on Computer Vision},
  year={2026}
}
```