使用 Adaboost 时访问和修改 OpenCV 决策树节点

Question

我正在从 30000 个随机生成的特征中学习一棵增强树。学习仅限于说最好的 100 个特征。在学习了如何从 CvBoost 对象中提取决策树使用的特征的索引之后。

我这样做的动机是消除生成所有 30000 个特征的要求，并且只计算将使用的那些特征。我已经包含了从 CvBoost.save 函数生成的 yml 文件的打印输出。我想我想要的是名为 sample_count 的值，它在深度 1 的决策树中标识如下所示的特征：

 trees:
      -
         best_tree_idx: -1
         nodes:
            -
               depth: 0
               sample_count: 11556
               value: -1.8339875099775065e+00
               norm_class_idx: 0
               Tn: 0
               complexity: 0
               alpha: 0.
               node_risk: 0.
               tree_risk: 0.
               tree_error: 0.
               splits:
                  - { var:497, quality:8.6223608255386353e-01,
                      le:5.3123302459716797e+00 }
            -
               depth: 1
               sample_count: 10702
               value: -1.8339875099775065e+00
               norm_class_idx: 0
               Tn: 0
               complexity: 0
               alpha: 0.
               node_risk: 0.
               tree_risk: 0.
               tree_error: 0.
            -
               depth: 1
               sample_count: 854
               value: 1.8339875099775065e+00
               norm_class_idx: 1
               Tn: 0
               complexity: 0
               alpha: 0.
               node_risk: 0.
               tree_risk: 0.
               tree_error: 0.

编辑

目前我有以下用于访问数据的代码：

//Interrogate the Decision Tree. Each element is a Decision Tree, making up the classifer
    CvSeq* decisionTree = boostDevice.get_weak_predictors();

    simplifyFeatureSet(decisionTree, firstOrderROIs );

这个函数是：

inline void Chnftrs::simplifyFeatureSet(CvSeq* decisionTree, std::vector<boost::tuple<int, cv::Rect> >& rois)
{
    //This variable stores the index of the feature used from rois and a pointer to the split so that the variable there can
    //be updated when the rois are pruned and reordered.
    std::vector<boost::tuple<int, CvDTreeSplit* > > featureIdx;

    //Determine the max depth of the tree

    printf("Size of boost %d \n", decisionTree->total);

    for (int i = 0; i < decisionTree->total; i++)
    {
            //Get the root of the tree
            CvBoostTree *tree =0;
            tree = (CvBoostTree*)cvGetSeqElem(decisionTree, i);

            if(tree == 0)
                printf("Tree is NULL\n");
            else
                printf("Tree Addr %ld\n", tree);            

            const CvDTreeNode *root = tree->get_root();

            printf("Class_idx %d, Value %f ", root->sample_count, root->value);

            featureIdx.push_back(boost::tuple<int, CvDTreeSplit*>(root->split->var_idx, root->split)); 

                    //Search down the right hand side
            depthFirstSearch(root->right, featureIdx);

            //Search down the left hand side
            depthFirstSearch(root->left, featureIdx);


    }
}

但是，当我尝试访问 root 的任何成员（例如 root->sample_count）时，我遇到了分段错误。除非 CvTreeTrainData.shared 设置为 true（默认情况下为 false），否则 CvTree 的成员可能无法访问。如图所示here

任何帮助都会很棒

干杯

彼得

Answer 1

好的，

能够按照源代码中关于 CvBoost 分类器如何从磁盘保存和读取自身的方法来编辑决策树。出于某种原因，cvGetSeqElem() 不会从传递给它的 CvSeq 对象返回有效指针，用于决策树类型对象。

为了获得决策树的副本，CvSeqReader 和宏 cvStartReadSeq 效果最好。宏 CV_READ_SEQ_ELEM() 似乎在获取 Seq 中的下一棵树的循环期间自我更新。：

    CvSeqReader reader;
    cvStartReadSeq( weak, &reader );

     for (int i = 0; i < weak->total; i++)
        {

            CvBoostTree* tree;
            CV_READ_SEQ_ELEM( tree, reader );

                const CvDTreeNode *root = 0;
                root = tree->get_root();

                printf("Root Split VarIdx : %d c: %f, ", root->split->var_idx, root->split->ord.c);

                featureIdx.push_back(boost::tuple<int, CvDTreeSplit*>(root->split->var_idx, root->split)); 

                //Search down the right hand side
                depthFirstSearch(root->right, featureIdx);

                //Search down the left hand side
                depthFirstSearch(root->left, featureIdx);


        }