Title: An Attribute-Based Measure of Video Complexity

Abstract:

This paper introduces VideoABC (Video Attribute-Based Complexity), a novel framework designed to estimate the difficulty that video-question pairs present to video-LLMs. In this context, video complexity is operationalized as the likelihood of a video-LLM failing to answer correctly for a specific video-question combination. VideoABC functions as a non-parametric metric that leverages a predefined vocabulary of attributes known to influence complexity—such as scene intricacy or the velocity of events relevant to the query—alongside a reference video dataset.

During the training phase, reference videos are mapped into the attribute space, which is subsequently quantized. The expected ABC value for each resulting quantization cell is then calculated. For new inputs, the complexity of a video and its corresponding question is determined by identifying the associated quantization cell and using its expected ABC value. To ensure the framework remains effective even when reference datasets are small, the method employs a dual-quantizer approach. This combines a k-means quantizer, which provides precise complexity estimates for samples within the reference dataset’s distribution, with a universal lattice quantizer that ensures robust generalization to out-of-distribution data.

To populate the cells of the lattice quantizer during training, the authors propose a synthetic video generation technique inspired by psychophysics studies involving target-distractor manipulations. This allows for the calculation of expected ABCs for these cells. Experimental findings demonstrate that VideoABC achieves high efficacy even with low-dimensional attribute representations, significantly surpassing methods like "video-LLM as judge" while requiring considerably less computational complexity. Furthermore, the study highlights that the explainability of VideoABC scores, derived from well-defined attributes, offers valuable insights into how the attribute composition of benchmarks influences their overall complexity.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC