Scholarlink

Visual Question Answering (VQA) is a challenging task that requires models to simultaneously process natural language and visual information to produce accurate answers. Transformer-based architectures have achieved considerable progress in this field, but they often fail to effectively utilize positional information, which is crucial for spatial reasoning. This study introduces the RPR-MCAoAN model, where relative positional information is integrated into attention mechanisms to improve spatial and semantic alignment between words and image regions. Experiments on the VQA-v2 dataset show that the model achieves overall accuracies of 71.70% on the test-dev split and 71.95% on the test-standard split. For different question categories, the model obtains 87.60% and 87.80% in Yes/No, 55.90% and 55.20% in Number, and 61.70% and 61.95% in Other, respectively. Additional ablation studies further confirm the effectiveness of incorporating relative positional information, demonstrating its importance in enhancing multimodal reasoning for VQA.

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

RPR-MCAoAN: A Transformer-Based Co-attention Network with Relative Positional Representations for Visual Question Answering

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pham Hoai Nhan,
  • Thai Gia Bao,
  • Nguyen Minh Hai

摘要

Visual Question Answering (VQA) is a challenging task that requires models to simultaneously process natural language and visual information to produce accurate answers. Transformer-based architectures have achieved considerable progress in this field, but they often fail to effectively utilize positional information, which is crucial for spatial reasoning. This study introduces the RPR-MCAoAN model, where relative positional information is integrated into attention mechanisms to improve spatial and semantic alignment between words and image regions. Experiments on the VQA-v2 dataset show that the model achieves overall accuracies of 71.70% on the test-dev split and 71.95% on the test-standard split. For different question categories, the model obtains 87.60% and 87.80% in Yes/No, 55.90% and 55.20% in Number, and 61.70% and 61.95% in Other, respectively. Additional ablation studies further confirm the effectiveness of incorporating relative positional information, demonstrating its importance in enhancing multimodal reasoning for VQA.