代码学习discriminative patch（一）

matlab code，对应论文：Unsupervised Discovery of Mid-Level Discriminative Patches Saurabh Singh, Abhinav Gupta, and Alexei A. Efros]

其中VisualEntityDetectors：是matlab中定义的类

train for Discriminative patches for pascal 2007 subset

train：
- getPascalData() 从voc2011 xml文件中获取图像annotation存入PASCAL_DATA.mat
- warpCrossValClusteringUnsup(1, USR, true);第三个参数在real train中应该是false,supposed to run on a cluster with shared file system
  - getdataset() load PASCAL_DATA.mat
  - getTrainValSplitForCategory(data, category)根据类别划分出正样本和负样本(负样本不包含category给出的类别)
  - processCategory()
    - getParamsForCategory(category) 获取所有可能用到的参数
    - getTrainValSplitUnsupervised(posData, negData)把正负样本各分一半 D1D2 N1N2用作交叉验证
    - processImages()
      - getRandomPatchesFromPyramid( ) 371x1984 742x1984
        
        sampleRandomPatches() 371x1984 742x1984
        
        convertToCanonicalSize() 将load的图像大小标准化
        
        getGradientImage()对标准化后的图像进行梯度计算
        
        constructFeaturePyramidForImg()计算19个不同scale下hog特征保存在pyramid中 19cell
        
        unentanglePyramid()依次将每个scale下的fea转换为1984=8x8x31维特征得到11220x1984
        
        通过去除重叠部分多没有梯度信息的patch最后得到正负(5/10)样本图像的patch371x1984 742x1984
      - clusterPatches() 正样本聚类得到95clusters(381/4)
      - refineClustersOverlap( )对聚类中心进行筛选得到42 剩余有效patch 193
        
        selectClustAboveThresh()聚类中心下的patch少于一定数目去除该cluster 57clusters
        
        去除每个聚类中心下patch中重复度大的patch 然后判断是否仍有足够数量的patch 决定是否留该聚类中心 42clusters
      - doTheIterations() 依次对D1之前每个cluster进行svmtrain 然后用D2的数据进行svmpredict 计算D2数据集下的得到的cluster下的patch(有某些限制) 并依次交换D1D2数据集进行重复运算最后得到42x1984的svmtrain model
        
        iterateTraining()

初次进行firstTrainDetectors() 如果存在直接load

constructVisDetFromModels(firstTrainDir, trainingData.selectedClusters, params, '_det');

hardNegMineTrainDetectors(instanceId, thisIterDir, hardNegOut,detectors, trainingData)

detectPresenceUsingDetectors(instanceId, thisIterDir, detectionOut,detectors, trainingData.selectedClusters)

getTopNDetsPerCluster(detectionResult, maxOverlap, trainingData.trainSetPos, numTopN)

prepareDetectedPatchClusters()

calculateClusterCenters()

Prepare the data for the next iterations

- assimilateCrossValBatchResults()

detect discriminative patches

load model: 全部训练数据用上得到的模型svm训练的模型总积 5481clusters 1984dimension features
load data and get data info: size
construct detection params: 初始化参数
detector.detectPresenceUsingEntDet(): 用model对该图像中patch属于哪个clusters计算score
displayPatchBox: 画出patch box
- constructFeaturePyramid() 19cell 不同尺度
- getDetectionsForEntDets ()
  - unentanglePyramid() 13011x1984 hog特征
  - Mysvmpredicted()在5841x1984model上计算出每个patch在cluster的score
  - SelectTopN()根据score判断对pathch的取舍