Related Experiment Video
Updated: Nov 28, 2025
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
9.4K
Feature-Aware Adaptation and Density Alignment for Crowd Counting in Video Surveillance
IEEE Transactions on Cybernetics
|December 1, 2020
Summary
This study introduces a novel domain adaptation method for crowd counting, effectively bridging the gap between synthetic and real-world data. The approach enhances model generalization for accurate crowd density estimation in unseen environments.
Area of Science:
- Computer Vision
- Deep Learning
- Artificial Intelligence
Background:
- Deep neural networks have advanced crowd counting and density estimation.
- Challenges include data scarcity for supervised learning and poor generalization to new domains.
Purpose of the Study:
- To propose a domain adaptation method for crowd counting to bridge the synthetic-to-real domain gap.
- To improve model performance on unseen real-world crowd scenes.
Main Methods:
- Developed a domain-adaptation-style crowd counting method.
- Implemented multilevel feature-aware adaptation (MFA) for domain-invariant features.
- Utilized structured density map alignment (SDA) for realistic density map generation.
Main Results:
- The proposed method effectively adapts models from synthetic to real-world scenes.
- Achieved superior performance on Shanghai Tech Part B, WorldExpo'10, Mall, and UCSD datasets.
- Outperformed state-of-the-art methods in cross-domain crowd counting.
Conclusions:
- The novel approach significantly reduces the domain gap in crowd counting.
- Demonstrates effective cross-domain generalization for crowd density estimation.