Tolerating Annotation Displacement in Dense Object Counting via Point Annotation Probability Map.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object counting in crowded scenes is a significant challenge for computer vision.
• Current deep learning methods often use Gaussian density regression, which may not properly handle annotation displacement from human annotators.

Purpose of the Study:

• To develop a more robust method for dense object counting by explicitly considering annotation displacement.
• To improve the accuracy and reliability of object counting in complex, crowded environments.

Main Methods:

• Proposed a generalized Gaussian distribution (GGD) to create a point annotation probability map (PAPM) for learning targets.
• Introduced a hand-designed PAPM (HD-PAPM) and an adaptively learned PAPM (AL-PAPM) to tolerate annotation displacement.
• Integrated the PAPM approach with existing methods, such as P2PNet, creating P2P-PAPM to enhance robustness.

Main Results:

• The proposed PAPM methods demonstrate superior robustness against annotation displacement compared to traditional approaches.
• Experiments show significant improvements in dense object counting accuracy when using the GGD-based PAPM.
• Integration with P2PNet (P2P-PAPM) also yielded enhanced robustness to annotation variations.

Conclusions:

• Considering annotation displacement is crucial for improving dense object counting performance.
• The GGD-based PAPM offers a flexible and effective solution for robust object counting.
• The proposed methods show promise for real-world applications requiring accurate crowd counting.