Related Experiment Video
Updated: Aug 5, 2026
08:25
Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment
Published on: May 7, 2019
Industrial Object Counting from Traditional Machine Vision to Open-World Foundation Models: A Systematic Review
Wei Wang1,2,3, Shengjie Zhang4, Jin He2
1School of Information and Communication Engineering, University of Electronic Science and Technology of China (UESTC), Chengdu 611731, China.
Sensors (Basel, Switzerland)
|July 28, 2026
Summary
This review traces the evolution of industrial object counting from traditional methods to deep learning, highlighting advancements in CNNs, Transformers, and foundation models for complex counting tasks. It proposes future directions like lightweight models and unsupervised learning for smarter manufacturing.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Machine Learning
Background:
- Industrial object counting is crucial for smart manufacturing, inventory management, and production monitoring.
- The field has transitioned from traditional machine vision to deep learning over the past 15 years.
- Key challenges include occlusion, object overlap, multi-scale variations, and dense object counting.
Purpose of the Study:
- To provide a comprehensive review of industrial object counting, focusing on technological evolution.
- To analyze the impact of deep learning on object counting tasks.
- To identify future research directions for more efficient and intelligent industrial object counting.
Main Methods:
- Review of traditional sensor-based and template-matching methods.
- Exploration of deep learning approaches including CNNs, Transformers, Mamba, and Large Foundation Models.
- Analysis of the paradigm shift towards Class-Agnostic Counting (CAC) and Exemplar-Free Counting.
Main Results:
- Deep learning models have significantly improved performance in handling complex counting scenarios.
- Class-Agnostic Counting (CAC) and Exemplar-Free Counting reduce data dependency and enhance generalization.
- Mainstream datasets (FSC-147, NWPU-MOC, OmniCount-191) and metrics (MAE, RMSE, PrACo) are systematically organized.
Conclusions:
- Current methods face challenges like high annotation costs, weak cross-domain adaptability, and real-time requirements.
- Future research should focus on lightweight models, unsupervised learning, multi-modal fusion, and prompt-based interactive counting.
- This review offers a technical blueprint for advancing industrial object counting.