Answers are grounded in this paper and cite their sources.
Demonstrates that adapting frozen DINOv3 improves crowd counting accuracy in variable scales, suggesting effective resolution conditioning.
Scale variation remains a fundamental challenge for intelligent surveillance sensors in crowd-counting and localization tasks. While recent selective inheritance methods have shown promise through multi-resolution feature fusion, they typically rely on conventional CNN backbones with limited representation capacity. Foundation models such as DINOv3 offer powerful self-supervised representations, yet directly applying frozen DINOv3 features to selective scale-aware counting is non-trivial: these features are semantically strong but not directly aligned with density estimation or scale-conditioned inheritance requirements. In this paper, we propose D3-CalibCount, a trainable-parameter-efficient framework for adapting frozen DINOv3 representations to selective scale-aware crowd counting and localization. We introduce a lightweight Scale Harmonization Adapter (SHA) that performs resolution-conditioned feature calibration, transforming generic DINOv3 representations into scale-selective counting features suitable for progressive inheritance across resolution levels. Extensive experiments on three widely used benchmarks show consistent improvements over selective inheritance baselines, especially under severe scale variation. From a deployment perspective, the method reduces the trainable portion of the network, while the frozen DINOv3-L backbone still introduces a higher inference cost than lighter CNN baselines. The target application scenario is camera-based intelligent surveillance sensing, where crowd density estimation is performed from visual sensor inputs at GPU-backed monitoring nodes. These results suggest that lightweight adaptation of frozen foundation features is a practical direction for crowd counting and other dense prediction tasks.
No takes yet. Share an insight, caveat, or question.
Xu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: