Automated crowd counting in Hajj video faces significant challenges, not due to the models' lack of capacity, but because the footage is shot from steep angles, individuals extensively occlude one another, and a single frame can contain over a thousand people. Existing benchmarks lack public access or detailed per-second data.
To address this, we revisit the HAJJv2 dataset and contribute HAJJv2-CrowdCount: per-second human-annotated crowd counts for its testing videos. Using these annotations, we benchmark three recent zero-shot counting paradigms: an open-vocabulary detector (YOLO-World), a point-based counter (APGCC), and a promptable segmentation-based counter (SAM3Count).
SAM3Count achieves the best overall mean absolute error (MAE 70.4, 95% CI 56.0-86.1), ahead of YOLO-World (92.0) and APGCC (152.9). However, in the deployment-relevant scenarios, the results reverse: on the densest frames, both detection and segmentation-based counters degrade sharply (MAE exceeding 300), while the point-based counter degrades more gracefully (MAE 114.9).
This inversion is crucial for Hajj crowd management, where reliable counts are most needed in the densest, most occluded scenes. The annotations are released to support the reproduction and extension of these results.
Blogger's Review: This study provides a new benchmark for crowd counting in Hajj, particularly highlighting the differences in applicability among various counting methods in dense and occluded environments. It serves as an important reference for future research and practical applications, especially as the demand for effective crowd management at large-scale events increases.