Decoupled Two-Stage Crowd Counting and Beyond.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Density map estimation is a common approach for counting dense objects like crowds.
• Existing methods struggle with ambiguous appearance cues in congested scenes, hindering individual identification and error diagnosis.
• Counting can be viewed as a two-stage process: identifying object regions and then determining exact counts.

Purpose of the Study:

• To introduce a novel decoupled two-stage counting (D2C) framework for improved object counting and localization.
• To address data deficiency and sample imbalance issues in object counting tasks.
• To enable reliable error analysis and enhance model generalizability.

Main Methods:

• Proposed a probabilistic intermediate representation called a probability map, indicating the likelihood of each pixel belonging to an object.
• Decoupled the counting process into probability map regression (PMR) and count map regression (CMR).
• Developed a D2C framework that sequentially regresses the probability map and then learns a counter conditioned on it, using a peak point detection algorithm for localization.

Main Results:

• The D2C framework demonstrated state-of-the-art counting and localization performance across six crowd counting benchmarks.
• Analysis revealed that the primary bottleneck in counting accuracy lies in Probability Map Regression (PMR).
• The proposed D2CNet exhibited remarkable cross-dataset transferability due to the appearance-independent nature of the probability map.

Conclusions:

• The D2C framework effectively improves object counting and localization by decoupling the process and utilizing probability maps.
• The probability map representation aids in addressing data deficiency and sample imbalance, leading to more reliable counters.
• D2CNet shows strong performance and generalizability, offering a robust solution for crowd counting and analysis.