Scholarlink

<p>Few-shot object counting aims to count an object of an arbitrary category using only a few annotated exemplars, i.e., few-shot. Existing methods have shown promising results in few-shot counting by using the transformer model combined with convolutions. However, these methods’ approaches to extracting the query and exemplars features using the separated feature extractors, namely Vision Transformer (ViT) and Convolution Neural Network (CNN), respectively, can cause the inconsistency of latent spaces. To address this issue, in this work, we proposed the ViT-CNN architecture sharing feature extraction for query and exemplar extraction by adapting the ViT-Adapter model as the backbone. Our proposed architecture also takes advantage of multi-scale features and the exemplar iterative interaction to enhance the result. Furthermore, we proposed a Density Map Refinement Module to refine the predicted density map using the exemplar information. Extensive experiments on the FSC-147 and CARPK datasets show that our approach significantly outperforms previous state-of-the-art approaches (7.20% error reduction <i>w.r.t.</i> test MAE and 6.35% <i>w.r.t.</i> test RMSE).</p>

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

Unified ViT-CNN for few-shot object counting

  • 期刊论文

  • 发表日期：

    2025年1月17日

查看全文
  • Khanh-An C. Quan,
  • Vinh-Tiep Nguyen,
  • Tam V. Nguyen,
  • Minh-Triet Tran

摘要

Few-shot object counting aims to count an object of an arbitrary category using only a few annotated exemplars, i.e., few-shot. Existing methods have shown promising results in few-shot counting by using the transformer model combined with convolutions. However, these methods’ approaches to extracting the query and exemplars features using the separated feature extractors, namely Vision Transformer (ViT) and Convolution Neural Network (CNN), respectively, can cause the inconsistency of latent spaces. To address this issue, in this work, we proposed the ViT-CNN architecture sharing feature extraction for query and exemplar extraction by adapting the ViT-Adapter model as the backbone. Our proposed architecture also takes advantage of multi-scale features and the exemplar iterative interaction to enhance the result. Furthermore, we proposed a Density Map Refinement Module to refine the predicted density map using the exemplar information. Extensive experiments on the FSC-147 and CARPK datasets show that our approach significantly outperforms previous state-of-the-art approaches (7.20% error reduction w.r.t. test MAE and 6.35% w.r.t. test RMSE).