VSI-Super-Wild: Benchmark for Spatial Supersensing in Long-Form Wild Videos

VSI-Super-Wild is a benchmark for evaluating multimodal models on spatial supersensing capabilities in long-form, in-the-wild videos. It was created by researchers from Tsinghua University, NVIDIA, and Stanford University for the ECCV 2026 conference. The dataset moves beyond short indoor clips and object-centric settings to study world state maintenance and prediction.

Use Cases

• Benchmarking spatial reasoning capabilities of multimodal models based on long-form video data.
• Training models for world state prediction based on continuous sensory streams described in the dataset.
• Evaluating AI performance on in-the-wild video understanding tasks beyond short indoor clips.

Strengths

• Created by a collaboration between Tsinghua University, NVIDIA, and Stanford University.
• Focuses on long-form, in-the-wild videos, moving beyond common short-clip benchmarks.
• Designed for the ECCV 2026 conference, indicating a research-oriented purpose.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and license information are unknown, which may limit suitability assessment.

Provenance

Source

THUSI-Lab, with collaboration from Tsinghua University, NVIDIA, and Stanford University.

Collection Method

Likely curated for research purposes, but specific gathering method is not detailed.

Freshness

Last updated 2026-07-14 15:29:48; freshness should be verified.