论文解读 Saliency Weighted Convolutional Features for Instance Search

论文解读 Saliency Weighted Convolutional Features for Instance Search

---

Saliency Weighted Convolutional Features for Instance Search，E Mohedano，K Mcguinness，X Giroinieto，arXiv.org

---

1 Contribution

• 1 We propose a novel approach to instance search combining saliency weighting over off-the-shelf convolutional features which are aggregated using a large vocabulary with a bag of words model.

• 2 We demonstrate that this weighting scheme gives outperforms all other state of the art on the challenging INSTRE benchmark.

• 3 higher performance on saliency benchmarks does not necessary equate to improved performance when used in the instance search task

    我对这篇的理解其实主要是利用显著性检测给BLCV过程赋一个权重，类似于文件检索里BOW中的tf-idf权重机制。

---

2 Pipeline

    在上图中，Smantic feature相当于BLCV的过程，详情查看上一篇博文。Saliency 相当于一个显著性检测网络，可以得到一张图片的显著性结果图，然后利用显著性结果图给BLCV过程中的到的Assignment map赋予权值，然后再统计词频生成最终的特征描述。

---

3 Saliency model

此图介绍了不同显著性检测model在MIT300的检测效果

此图可视化了显著性检测model的结果

---

4 Experiments

• 1 Aggregation methods
  

  
  

• 2 Comparison with the state-of-the-art
  

  
  
  不加查询扩展的实验对比结果
  

  
  
  加了查询扩展的实验对比结果

• 3 结果展示