Application of improved transformer based on weakly supervised in crowd localization and crowd counting

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current crowd localization methods involve complex pre-processing and post-processing steps.
• Existing techniques often rely on pseudo bounding boxes and pre-designed positioning maps, limiting efficiency.

Purpose of the Study:

• To propose an end-to-end crowd localization framework, WSITrans, addressing the limitations of current methods.
• To reformulate weakly-supervised crowd localization using Transformer architecture for improved crowd counting.

Main Methods:

• Implemented a pure Transformer architecture with global maximum pooling (GMP) after each stage to enhance head detail extraction.
• Designed a binarization module to refine decoder output features and fuse confidence scores for greater accuracy.

Main Results:

• WSITrans achieved significant improvements on three challenging crowd localization benchmarks.
• The proposed method enhanced the F1-measure by 4.0%, demonstrating superior performance.

Conclusions:

• WSITrans offers an effective end-to-end solution for weakly-supervised crowd localization and counting.
• The framework simplifies the process while achieving state-of-the-art results in crowd analysis.