VSI-Super-Wild

Towards Spatial Supersensing in the Wild
ECCV 2026

VSI-Super-Wild is a long-form video benchmark for evaluating spatial supersensing in the wild. It extends spatial reasoning evaluation from object-centric indoor settings to real-world, long-horizon videos that require models to build and query world states over time.

The benchmark contains view-level videos, human-verified question-answer pairs, and task prompts covering agent-centric motion, environment/place ordering, object ordering, and object counting.

Dataset Summary

Item	Value
View-level videos	442
Source videos / nicknames	113
Total duration	284.52 hours
Scene categories	8
Average video duration	38.62 minutes
Minimum video duration	0.87 minutes
Maximum video duration	261.08 minutes
Human-verified Q&A pairs	6,980

Tasks

The benchmark includes four task families:

Task	Description	Answer format	Count
VMR	Motion orientation recall: infer the camera/person motion direction at a queried moment.	Multiple choice	1,215
VPO	Place temporal ordering: order queried places or frames by capture time.	Multiple choice	1,302
VOO	Object temporal ordering: order objects by first or last visible occurrence.	Multiple choice	3,350
VOC	Continuous object counting: count unique instances of a target object category.	Integer	1,113

File Organization

The repository is organized as:

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

The video archives are grouped by video duration bucket in minutes:

Duration bucket	Files	Approx. size
videos_0_10_part_*.zip	1	49.53 GB
videos_10_30_part_*.zip	2	267.49 GB
videos_30_60_part_*.zip	6	833.35 GB
videos_60_120_part_*.zip	2	265.15 GB
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 video file used by each question.

Data Usage

VSI-Super-Wild is intended for research on long-video understanding, spatial reasoning, world-state modeling, and multimodal model evaluation. The benchmark is designed to test whether models can construct, maintain, update, and query spatial world states over long, in-the-wild video streams.

After downloading and extracting the video archives, place the extracted videos under the repository-level data/ directory of the GitHub codebase, or set VSI_SUPER_WILD_VIDEO_ROOT to your extracted video path. The evaluation scripts and task definitions are maintained in the GitHub repository.

Expected evaluation layout:

VSI-Super-Wild/
|-- data/
|   |-- long_video_persp/
|   |-- new_long_video_persp/
|   `-- top20merge_0207_persp/
`-- tasks/vsi_super_wild/data/vsi_super_wild_qa.jsonl

Citation

If you use VSI-Super-Wild, please cite:

@misc{vsi_super_wild_2026,
  title        = {Towards Spatial Supersensing in the Wild},
  author       = {Gu, Tianjun and Xin, Tianyu and Zhang, Kuan and Yang, Bowen and Chua, Kok-Chung and Zhang, Xinran and Li, Peize and Chen, Yupeng and Zhao, Qiyue and Xie, Qinlei and Liu, Jianhang and Lu, Yucheng and Han, Yinan and Pavone, Marco and Li, Yiming},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/THUSI-Lab/VSI-Super-Wild}}
}