SpookyBench: A Benchmark for Purely Temporal Video Understanding

SpookyBench is a novel benchmark dataset designed to evaluate the ability of video-language models (VLMs) to understand purely temporal patterns, independent of spatial cues. The dataset consists of 451 videos across four categories: Text, Object Images, Dynamic Scenes, and Shapes. Each video appears as random noise in individual frames, but reveals meaningful content (words, objects, etc.) when viewed as a temporal sequence. This design exposes a critical limitation in current VLMs, which often heavily rely on spatial information and struggle to extract meaning from purely temporal sequences.

Paper: Time Blindness: Why Video-Language Models Can't See What Humans Can?

Project Website: https://timeblindness.github.io/

The dataset contains 451 videos distributed as follows:

Category	Total Videos	Description
Text	210 (46.6%)	English words encoded through temporal noise patterns
Object Images	156 (34.6%)	Single objects encoded using temporal animation
Dynamic Scenes	57 (12.6%)	Video depth maps with temporal motion patterns
Shapes	28 (6.2%)	Geometric patterns encoded through temporal sequences
Total	451	Comprehensive temporal understanding evaluation

Download: You can download the dataset from Hugging Face using the following command:

wget https://huggingface.co/datasets/timeblindness/spooky-bench/resolve/main/spooky_bench.zip
unzip spooky_bench.zip

License: MIT License