Scholarlink

We propose a multi-task deep learning framework for pedestrian attribute classification, jointly predicting five key attributes: upper and lower clothing colors, gender, and the presence of a bag and hat. Our approach employs a ConvNeXt-Large backbone pretrained on ImageNet-22k, enhanced with advanced data augmentation techniques including MixUp and CutMix. To address class imbalance, we introduce a conditional augmentation strategy targeting challenging color classes (e.g., black and gray). Experiments on the MIVIA PAR KD Dataset 2025, comprising over 105,000 annotated pedestrian images, demonstrate the effectiveness of our method, achieving a peak mean accuracy of 93.14% across all attributes. These results highlight the advantages of combining modern convolutional architectures, task-aware augmentation, and multi-task learning for robust pedestrian analysis in real-world settings.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-task Pedestrian Attribute Classification Using ConvNeXt with Advanced Data Augmentation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Magzhan Kairanbay,
  • Ali Salman

摘要

We propose a multi-task deep learning framework for pedestrian attribute classification, jointly predicting five key attributes: upper and lower clothing colors, gender, and the presence of a bag and hat. Our approach employs a ConvNeXt-Large backbone pretrained on ImageNet-22k, enhanced with advanced data augmentation techniques including MixUp and CutMix. To address class imbalance, we introduce a conditional augmentation strategy targeting challenging color classes (e.g., black and gray). Experiments on the MIVIA PAR KD Dataset 2025, comprising over 105,000 annotated pedestrian images, demonstrate the effectiveness of our method, achieving a peak mean accuracy of 93.14% across all attributes. These results highlight the advantages of combining modern convolutional architectures, task-aware augmentation, and multi-task learning for robust pedestrian analysis in real-world settings.