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

2026-07-08 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Advanced, quick

Summary

HAJJv2-CrowdCount is a new per-second human-annotated crowd counting benchmark for the HAJJv2 dataset, designed to address the challenges of dense crowd footage from steep, near-vertical camera angles with extensive occlusion and over a thousand people per frame. This benchmark evaluates three zero-shot counting paradigms: YOLO-World (an open-vocabulary detector), APGCC (a point-based counter), and SAM3Count (a promptable segmentation-based counter). SAM3Count achieved the lowest overall mean absolute error (MAE 70.4, 95% CI 56.0-86.1), outperforming YOLO-World (92.0) and APGCC (152.9). However, in the densest frames, crucial for Hajj crowd management, this order reverses; detection- and segmentation-based counters degrade sharply (MAE exceeding 300), while the point-based APGCC degrades more gracefully (MAE 114.9). The annotations are publicly released.

Key takeaway

For Computer Vision Engineers developing crowd management systems, especially in high-density, occluded environments like Hajj, your model selection should prioritize performance on the densest frames over overall mean absolute error. While segmentation-based models might show lower average MAE, point-based counters like APGCC demonstrate superior robustness and reliability when counts are most critically needed in extremely dense scenes, preventing sharp degradation.

Key insights

Zero-shot crowd counting performance in dense, occluded scenes reverses model efficacy compared to overall metrics.

Principles

• Dense crowd counting requires specialized model robustness.
• Overall MAE may not reflect real-world deployment needs.

Method

The method involves annotating per-second crowd counts for HAJJv2 test videos and benchmarking zero-shot models (YOLO-World, APGCC, SAM3Count) against these annotations, specifically analyzing performance in dense frames.

In practice

• Evaluate crowd counting models on scene density subsets.
• Consider point-based counters for extremely dense environments.

Topics

• Crowd Counting
• Zero-Shot Learning
• HAJJv2-CrowdCount
• Dense Scenes
• Computer Vision Benchmarks
• APGCC

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Never Seen Before: Benchmarking Genuine Zero-Shot Composed Image Retrieval with Consistent Video-Sourced Datasets — ZeroSight provides a true zero-shot CIR benchmark and a training-free MLLM method, revealing inflated existing performance.
• Multi-Modal Hyper-Graph Fusion for Low-Light Crowd Counting — Multi-modal hyper-graph fusion significantly improves low-light crowd counting by integrating RGB, depth, and edge cues.
• Molmo 2 | Counting objects and actions — Momo 2 excels in video object and action counting/pointing, while Gemini 3 shows stronger temporal understanding.
• Does it Really Count? Assessing Semantic Grounding in Text-Guided Class-Agnostic Counting — Current class-agnostic counting models struggle with semantic grounding, leading to unreliable object class identification from text prompts.
• YOLOv26: An Object Detector Built for Real-Time Deployment — YOLOv26 prioritizes edge deployment through architectural simplification and hardware-friendly optimizations.
• DF3DV-1K: A Large-Scale Dataset and Benchmark for Distractor-Free Novel View Synthesis — DF3DV-1K is a new dataset for benchmarking distractor-free novel view synthesis methods.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.