Back to Feed
Multimodal / Benchmarks & Evals

Testing Multimodal Models on Real-World Video

Original: Towards Spatial Supersensing in the Wild

Listen to the summary

Uses a voice available on your device

Audio options
On this page 4 sections
Related concepts 5 concepts

Key Takeaways

  • Researchers developed VSI-Super-Wild, a dataset consisting of 442 unconstrained panoramic YouTube videos across 8 scene categories.
  • The benchmark introduces four tasks: motion orientation recall, place temporal ordering, object temporal ordering, and continuous object counting.
  • Gemini-3.1-Pro led all models with a benchmark score of 44.36.
  • Open-source models show a performance decline when moving from short videos to those exceeding 120 minutes, dropping from 35.0 to 26.3.

Summary & Methodology Analysis

The VSI-Super-Wild framework processes raw panoramic video by projecting frames into four orthogonal perspective views. To generate ground truth metadata for the benchmark, the authors used a pipeline involving human experts and multimodal large language models (MLLMs) to identify objects. They automated object detection using YOLO-World and instance masking via Segment-Anything-3 (SAM3), while leveraging Depth-Anything-3 (DA3) to estimate camera poses. This allows for cognitively-grounded evaluation across complex, real-world scenes without relying on synthetic editing or constrained environments.

Interactive System Flowchart

Click diagram to expand and zoom

Cross-Examination & FAQs

A deeper dive clarifying mechanics, constraints, and baseline evaluations.

Q1. What is the primary goal of this research?

The goal is to move beyond synthetic benchmarks and static image understanding to evaluate how models construct and maintain a 3D world model over long-horizon, in-the-wild video.

Q2. What kind of data does the benchmark use?

The benchmark uses 442 unconstrained, panoramic YouTube videos spanning 8 categories like street view, retail, and office education.

Q3. Which model performed the best?

Gemini-3.1-Pro achieved the highest overall score of 44.36 on the benchmark.

Q4. How are the evaluation tasks constructed?

Tasks include motion orientation recall, place temporal ordering, object temporal ordering, and continuous object counting, all verified through a human-in-the-loop process.

Q5. How does video duration affect model performance?

Open-source models show a significant drop in average performance from 35.0 for videos under 10 minutes to 26.3 for videos exceeding 120 minutes.

Q6. What mechanism is used to avoid option-order bias?

The authors employ a 'Shuffle24' protocol during the evaluation of multiple-choice questions.

Q7. What is a major limitation of current models regarding memory?

Models perform worse on last-occurrence queries compared to first-occurrence queries in object temporal ordering tasks, suggesting they struggle to update their internal representations as new temporal evidence arrives.

Q8. What tools were used for spatial and temporal metadata generation?

The pipeline utilizes YOLO-World for detection, Segment-Anything-3 for instance masking, and Depth-Anything-3 for camera pose estimation.

Q9. Are there specific computational costs or training requirements mentioned?

The paper does not specify computational costs, latency figures, or training requirements.