当前位置：Gxlcms > 数据库问题 > 开源词袋模型DBow3原理&源码

开源词袋模型DBow3原理&源码

时间：2021-07-01 10:21:17 帮助过：13人阅读

vocabulary: 2 k: 9 //聚类中心数 3 L: 3 //层数 4 scoringType: 0 5 weightingType: 0 6 nodes: 7 - { nodeId:19, parentId:10, weight:2.8768207245178085e-01, 8 descriptor:dbw3 0 32 124 169 185 96 221 205 85 157 235 189 172 8 159 181 72 50 137 222 236 88 26 107 250 49 251 221 21 127 210 106 198 42 } 9 - { nodeId:20, parentId:10, weight:1.3862943611198906e+00, 10 descriptor:dbw3 0 32 120 164 24 104 249 61 80 95 115 27 172 24 31 147 64 248 145 152 76 56 26 111 82 23 219 223 17 125 226 200 202 42 } 11 - { nodeId:21, parentId:10, weight:0., 12 descriptor:dbw3 0 32 124 165 248 102 209 109 83 25 99 173 173 8 115 241 72 50 16 222 68 56 26 114 248 2 255 205 37 121 226 234 198 34 } 13 - { nodeId:22, parentId:10, weight:0., 14 descriptor:dbw3 0 32 36 161 252 81 224 233 97 139 235 53 108 40 151 199 204 26 3 158 110 24 18 248 234 65 205 152 53 117 194 200 198 162 } 15 - { nodeId:23, parentId:10, weight:1.3862943611198906e+00, 16 descriptor:dbw3 0 32 76 140 56 218 221 253 113 218 95 53 44 24 24 211 106 116 81 154 18 24 18 26 122 245 95 29 17 109 146 137 212 106 } 17 words: 18 - { wordId:0, nodeId:19 } 19 - { wordId:1, nodeId:20 } 20 - { wordId:2, nodeId:21 } 21 - { wordId:3, nodeId:22 } 22 - { wordId:4, nodeId:23 } 23 - { wordId:5, nodeId:24 }

是对聚类生成的树的描述，树上的每一个点都是node，只有叶子结点才是word，每张图进来提取特征，利用descriptor算特征距离，最终落到叶子上（单词），所有特征的单词构成该图片的词向量。

技术分享图片

检索入口描述：

  1. <span style="color: #008080"> 1</span> <span style="color: #000000">database:
  2. </span><span style="color: #008080"> 2</span> <span style="color: #000000">   nEntries: 4
  3. </span><span style="color: #008080"> 3</span> <span style="color: #000000">   usingDI: 0
  4. </span><span style="color: #008080"> 4</span> <span style="color: #000000">   diLevels: 0
  5. </span><span style="color: #008080"> 5</span> <span style="color: #000000">   invertedIndex:
  6. </span><span style="color: #008080"> 6</span> <span style="color: #000000">      -  <span style="color: #008000">//（wordId:0）
  7. </span></span><span style="color: #008080"> 7</span> <span style="color: #000000">         - { imageId:1, weight:1.6807896319101980e-03 }
  8. </span><span style="color: #008080"> 8</span> <span style="color: #000000">         - { imageId:2, weight:3.2497152852064880e-03 }
  9. </span><span style="color: #008080"> 9</span> <span style="color: #000000">         - { imageId:3, weight:3.6665308718065778e-03 }
  10. </span><span style="color: #008080">10</span> <span style="color: #000000">      -  <span style="color: #000000"><span style="color: #008000">//（wordId:1）</span></span>
  11. </span><span style="color: #008080">11</span> <span style="color: #000000">         - { imageId:1, weight:4.0497295661974788e-03 }
  12. </span><span style="color: #008080">12</span> <span style="color: #000000">      -  <span style="color: #000000"><span style="color: #008000">//（wordId:2）</span></span>
  13. </span><span style="color: #008080">13</span> <span style="color: #000000">         []
  14. </span><span style="color: #008080">14</span> <span style="color: #000000">      -
  15. </span><span style="color: #008080">15</span> <span style="color: #000000">         []
  16. </span><span style="color: #008080">16</span> <span style="color: #000000">      -
  17. </span><span style="color: #008080">17</span> <span style="color: #000000">         - { imageId:2, weight:3.9149658655580431e-03 }
  18. </span><span style="color: #008080">18</span> <span style="color: #000000">      -
  19. </span><span style="color: #008080">19</span> <span style="color: #000000">         - { imageId:3, weight:4.4171079458813099e-03 }
  20. </span><span style="color: #008080">20</span> <span style="color: #000000">      -
  21. </span><span style="color: #008080">21</span> <span style="color: #000000">         - { imageId:1, weight:2.0248647830987394e-03 }
  22. </span><span style="color: #008080">22</span>          - { imageId:3, weight:4.4171079458813099e-03 }

检索入口：

技术分享图片技术分享图片

根据voc字典里的描述，word id 2对应node id 21, 而node id 21对应的权值为0，也就是说word 2太普通了，在用来生成视觉词汇表的4张图里都出现了（参考中文文章里的“的”、“在”、“和”等常见词），不具有代表性， 于是根本就没有对应入口id，这是合理的。

 

开源出来的代码不是对相同word的入口进行加1投票，而是直接计算单词对应的所有EntryId分数，最后排序取前n个。分数可以有L1 L2 KL等几种计算方式

queryL1，C++不熟看了半天，用到map函数，注释：

  1. <span style="color: #008080"> 1</span> <span style="color: #0000ff">void</span> Database::queryL1(<span style="color: #0000ff">const</span> BowVector &<span style="color: #000000">vec, </span>QueryResults &ret, <span style="color: #0000ff">int</span> max_results, <span style="color: #0000ff">int</span> max_id) <span style="color: #0000ff">const</span>
  2. <span style="color: #008080"> 3</span> <span style="color: #000000">{
  3. </span><span style="color: #008080"> 4</span> <span style="color: #000000">  BowVector::const_iterator vit;
  4. </span><span style="color: #008080"> 5</span> 
  5. <span style="color: #008080"> 6</span>   std::map<EntryId, <span style="color: #0000ff">double</span>><span style="color: #000000"> pairs;
  6. </span><span style="color: #008080"> 7</span>   std::map<EntryId, <span style="color: #0000ff">double</span>><span style="color: #000000">::iterator pit;
  7. </span><span style="color: #008080"> 8</span> 
  8. <span style="color: #008080"> 9</span>   <span style="color: #0000ff">for</span>(vit = vec.begin(); vit != vec.end(); ++<span style="color: #000000">vit)
  9. </span><span style="color: #008080">10</span> <span style="color: #000000">  {
  10. </span><span style="color: #008080">11</span>     <span style="color: #0000ff">const</span> WordId word_id = vit-><span style="color: #000000">first;
  11. </span><span style="color: #008080">12</span>     <span style="color: #0000ff">const</span> WordValue& qvalue = vit-><span style="color: #000000">second;
  12. </span><span style="color: #008080">13</span> 
  13. <span style="color: #008080">14</span>     <span style="color: #0000ff">const</span> IFRow& row =<span style="color: #000000"> m_ifile[word_id];
  14. </span><span style="color: #008080">15</span> 
  15. <span style="color: #008080">16</span>     <span style="color: #008000">//</span><span style="color: #008000"> IFRows are sorted in ascending entry_id order</span>
  16. <span style="color: #008080">18</span>     <span style="color: #0000ff">for</span>(auto rit = row.begin(); rit != row.end(); ++<span style="color: #000000">rit)
  17. </span><span style="color: #008080">19</span> <span style="color: #000000">    {
  18. </span><span style="color: #008080">20</span>       <span style="color: #0000ff">const</span> EntryId entry_id = rit-><span style="color: #000000">entry_id;
  19. </span><span style="color: #008080">21</span>       <span style="color: #0000ff">const</span> WordValue& dvalue = rit-><span style="color: #000000">word_weight;
  20. </span><span style="color: #008080">22</span> 
  21. <span style="color: #008080">23</span>       <span style="color: #0000ff">if</span>((<span style="color: #0000ff">int</span>)entry_id < max_id || max_id == -<span style="color: #800080">1</span><span style="color: #000000">)
  22. </span><span style="color: #008080">24</span> <span style="color: #000000">      {
  23. </span><span style="color: #008080">25</span>         <span style="color: #0000ff">double</span> value = fabs(qvalue - dvalue) - fabs(qvalue) -<span style="color: #000000"> fabs(dvalue);
  24. </span><span style="color: #008080">26</span> 
  25. <span style="color: #008080">27</span>         pit =<span style="color: #000000"> pairs.lower_bound(entry_id);
  26. </span><span style="color: #008080">28</span>         <span style="color: #0000ff">if</span>(pit != pairs.end() && !(pairs.key_comp()(entry_id, pit-><span style="color: #000000">first)))
  27. </span><span style="color: #008080">29</span>         {  
  28. <span style="color: #008080">30</span>           pit->second +=<span style="color: #000000"> value;  <span style="color: #008000">//<span style="color: #008000">如果已经有entry_id,累加和</span></span>
  29. </span><span style="color: #008080">31</span> <span style="color: #000000">        }
  30. </span><span style="color: #008080">32</span>         <span style="color: #0000ff">else</span>
  31. <span style="color: #008080">33</span>         {                        <span style="color: #008000">//</span><span style="color: #008000">如果没有，插入此id</span>
  32. <span style="color: #008080">34</span>           pairs.insert(pit, std::map<EntryId, <span style="color: #0000ff">double</span>><span style="color: #000000">::value_type(entry_id, value));
  33. </span><span style="color: #008080">35</span> <span style="color: #000000">        }
  34. </span><span style="color: #008080">36</span> <span style="color: #000000">      }</span>
  35. <span style="color: #008080">38</span>     } <span style="color: #008000">//</span><span style="color: #008000"> for each inverted row</span>
  36. <span style="color: #008080">39</span>   } <span style="color: #008000">//</span><span style="color: #008000"> for each query word
  37. </span><span style="color: #008080">40</span> 
  38. <span style="color: #008080">41</span>   <span style="color: #008000">//</span><span style="color: #008000"> move to vector</span>
  39. <span style="color: #008080">42</span> <span style="color: #000000">  ret.reserve(pairs.size());
  40. </span><span style="color: #008080">43</span>   <span style="color: #0000ff">for</span>(pit = pairs.begin(); pit != pairs.end(); ++<span style="color: #000000">pit)
  41. </span><span style="color: #008080">44</span> <span style="color: #000000">  {
  42. </span><span style="color: #008080">45</span>     ret.push_back(Result(pit->first, pit-><span style="color: #000000">second));
  43. </span><span style="color: #008080">46</span> <span style="color: #000000">  }
  44. </span><span style="color: #008080">47</span> 
  45. <span style="color: #008080">48</span>   <span style="color: #008000">//</span><span style="color: #008000"> resulting "scores" are now in [-2 best .. 0 worst]</span>
  46. <span style="color: #008080">50</span>   <span style="color: #008000">//</span><span style="color: #008000"> sort vector in ascending order of score</span>
  47. <span style="color: #008080">51</span> <span style="color: #000000">  std::sort(ret.begin(), ret.end());
  48. </span><span style="color: #008080">52</span>   <span style="color: #008000">//</span><span style="color: #008000"> (ret is inverted now --the lower the better--)</span>
  49. <span style="color: #008080">54</span>   <span style="color: #008000">//</span><span style="color: #008000"> cut vector</span>
  50. <span style="color: #008080">55</span>   <span style="color: #0000ff">if</span>(max_results > <span style="color: #800080">0</span> && (<span style="color: #0000ff">int</span>)ret.size() ><span style="color: #000000"> max_results)
  51. </span><span style="color: #008080">56</span> <span style="color: #000000">    ret.resize(max_results);
  52. </span><span style="color: #008080">57</span> 
  53. <span style="color: #008080">58</span>   <span style="color: #008000">//</span><span style="color: #008000"> complete and scale score to [0 worst .. 1 best]
  54. </span><span style="color: #008080">59</span>   <span style="color: #008000">//</span><span style="color: #008000"> ||v - w||_{L1} = 2 + Sum(|v_i - w_i| - |v_i| - |w_i|)
  55. </span><span style="color: #008080">60</span>   <span style="color: #008000">//</span><span style="color: #008000">        for all i | v_i != 0 and w_i != 0</span><span style="color: #008080">62</span>   <span style="color: #008000">//</span><span style="color: #008000"> scaled_||v - w||_{L1} = 1 - 0.5 * ||v - w||_{L1}</span>
  56. <span style="color: #008080">63</span> <span style="color: #000000">  QueryResults::iterator qit;
  57. </span><span style="color: #008080">64</span>   <span style="color: #0000ff">for</span>(qit = ret.begin(); qit != ret.end(); qit++<span style="color: #000000">)
  58. </span><span style="color: #008080">65</span>     qit->Score = -qit->Score/<span style="color: #800080">2.0</span><span style="color: #000000">;
  59. </span><span style="color: #008080">66</span> }

 

开源词袋模型DBow3原理&源码

标签：gty   ons   忽略   数据   and   key   list   函数   div   

人气教程排行

本站所有资源全部来源于网络，若本站发布的内容侵害到您的隐私或者利益，请联系我们删除！