Aerial and remote sensing video is increasingly evaluated the same way as web video: multimodal large language models are shown sampled frames and asked multiple-choice questions, then scored on accuracy. A new meta-analysis on arXiv argues that this practice is fundamentally mismatched with the domain.
The paper's title points to the core claim: the gap is in the geometry. Standard video benchmarks reward recognition of objects and actions, but aerial footage demands reasoning about scale, perspective, motion, and spatial relationships that sampled frames and text questions rarely capture. By treating aerial video as just another web video, the field may be measuring the wrong capability.
The authors call for a convergent approach that centers geometric understanding in evaluation. While the abstract does not provide specific alternative metrics, the direction is clear: benchmarks for remote sensing video should be built around the structural properties that make the domain distinct, not borrowed from generic video QA. The analysis is a critique of current practice and a proposal to refocus how progress is measured.