HAJJv2-CrowdCount: Zero-Shot Benchmark for Dense Crowd Counting

The paper revisits the HAJJv2 dataset and releases per-second human crowd-count annotations for its test videos. It benchmarks three zero-shot paradigms: YOLO-World, APGCC, and SAM3Count. SAM3Count achieves the best overall mean absolute error, 70.4, with a reported 95% confidence interval of 56.0–86.1, followed by YOLO-World at 92.0 and APGCC at 152.9. However, performance changes in the densest frames: detection- and segmentation-based methods exceed 300 MAE, while the point-based APGCC degrades more gradually to 114.9 MAE.