Answers are grounded in this paper and cite their sources.
Humans consistently outperform state-of-the-art AI models in action recognition, particularly in challenging real-world conditions involving low resolution, occlusion, and visual clutter.e those utilised by humans Understanding the sources of this performance gap is essential for developing more robust and human-aligned models. In this paper, we present a large-scale human AI comparative study of egocentric action recognition using human-defined Minimal Identifiable Recognition Crops (MIRCs), defined as the smallest spatial or spatiotemporal regions sufficient for reliable human recognition. We used our previously introduced, Epic-ReduAct, a systematically spatially reduced and temporally scrambled dataset derived from 36 EPIC-KITCHENS videos, spanning multiple spatial reduction levels and temporal conditions. Recognition performance is evaluated using over 3,000 human participants, Side4Video, and Qwen 3.5 9B models. Our analysis combines quantitative metrics, Average Reduction Rate and Recognition Gap, with qualitative analyses of spatial (high-, mid-, and low-level visual features) and spatiotemporal factors, including a categorisation of actions into Low Temporal Actions (LTA) and High Temporal Actions (HTA). Results show that human performance exhibits sharp declines when transitioning from MIRCs to sub-MIRCs, reflecting a strong reliance on sparse, semantically critical cues such as hand object interactions. In contrast, the model degrades more gradually and often relies on contextual and mid- to low-level features, sometimes even exhibiting increased confidence under spatial reduction. Temporally, humans remain robust to scrambling when key spatial cues are preserved, whereas the model often shows insensitivity to temporal disruption, revealing class-dependent temporal sensitivities. Overall, our findings demonstrate that strong benchmark performance can obscure fundamental human AI misalignment and highlight limitations in current spatial and spatiotemporal representations, suggesting concrete directions for more efficient and human-aligned action recognition systems.
No takes yet. Share an insight, caveat, or question.
Rahmaniboldaji et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: