Merge remote-tracking branch 'upstream/3.4' into merge-3.4

modules/ml/src/boost.cpp

@@ -490,6 +490,7 @@ public:
 
     float predict( InputArray samples, OutputArray results, int flags ) const CV_OVERRIDE
     {
+        CV_CheckEQ(samples.cols(), getVarCount(), "");
         return impl.predict(samples, results, flags);
     }
 

modules/ml/src/rtrees.cpp

@@ -480,6 +480,7 @@ public:
     float predict( InputArray samples, OutputArray results, int flags ) const CV_OVERRIDE
     {
         CV_TRACE_FUNCTION();
+        CV_CheckEQ(samples.cols(), getVarCount(), "");
         return impl.predict(samples, results, flags);
     }
 

modules/ml/src/tree.cpp

@@ -43,6 +43,8 @@
 #include "precomp.hpp"
 #include <ctype.h>
 
+#include <opencv2/core/utils/logger.hpp>
+
 namespace cv {
 namespace ml {
 
@@ -1694,11 +1696,14 @@ void DTreesImpl::write( FileStorage& fs ) const
 void DTreesImpl::readParams( const FileNode& fn )
 {
     _isClassifier = (int)fn["is_classifier"] != 0;
-    /*int var_all = (int)fn["var_all"];
-    int var_count = (int)fn["var_count"];
-    int cat_var_count = (int)fn["cat_var_count"];
+    int varAll = (int)fn["var_all"];
+    int varCount = (int)fn["var_count"];
+    /*int cat_var_count = (int)fn["cat_var_count"];
     int ord_var_count = (int)fn["ord_var_count"];*/
 
+    if (varAll <= 0)
+        CV_Error(Error::StsParseError, "The field \"var_all\" of DTree classifier is missing or non-positive");
+
     FileNode tparams_node = fn["training_params"];
 
     TreeParams params0 = TreeParams();
@@ -1723,11 +1728,38 @@ void DTreesImpl::readParams( const FileNode& fn )
     readVectorOrMat(fn["var_idx"], varIdx);
     fn["var_type"] >> varType;
 
-    int format = 0;
-    fn["format"] >> format;
-    bool isLegacy = format < 3;
+    bool isLegacy = false;
+    if (fn["format"].empty())  // Export bug until OpenCV 3.2: https://github.com/opencv/opencv/pull/6314
+    {
+        if (!fn["cat_ofs"].empty())
+            isLegacy = false;  // 2.4 doesn't store "cat_ofs"
+        else if (!fn["missing_subst"].empty())
+            isLegacy = false;  // 2.4 doesn't store "missing_subst"
+        else if (!fn["class_labels"].empty())
+            isLegacy = false;  // 2.4 doesn't store "class_labels"
+        else if ((int)varType.size() != varAll)
+            isLegacy = true;  // 3.0+: https://github.com/opencv/opencv/blame/3.0.0/modules/ml/src/tree.cpp#L1576
+        else if (/*(int)varType.size() == varAll &&*/ varCount == varAll)
+            isLegacy = true;
+        else
+        {
+            // 3.0+:
+            // - https://github.com/opencv/opencv/blame/3.0.0/modules/ml/src/tree.cpp#L1552-L1553
+            // - https://github.com/opencv/opencv/blame/3.0.0/modules/ml/src/precomp.hpp#L296
+            isLegacy = !(varCount + 1 == varAll);
+        }
+        CV_LOG_INFO(NULL, "ML/DTrees: possible missing 'format' field due to bug of OpenCV export implementation. "
+                "Details: https://github.com/opencv/opencv/issues/5412. Consider re-exporting of saved ML model. "
+                "isLegacy = " << isLegacy);
+    }
+    else
+    {
+        int format = 0;
+        fn["format"] >> format;
+        CV_CheckGT(format, 0, "");
+        isLegacy = format < 3;
+    }
 
-    int varAll = (int)fn["var_all"];
     if (isLegacy && (int)varType.size() <= varAll)
     {
         std::vector<uchar> extendedTypes(varAll + 1, 0);

modules/ml/test/test_rtrees.cpp

@@ -95,6 +95,25 @@ TEST(ML_RTrees, 11142_sample_weights_classification)
     EXPECT_GE(error_with_weights, error_without_weights);
 }
 
+TEST(ML_RTrees, bug_12974_throw_exception_when_predict_different_feature_count)
+{
+    int numFeatures = 5;
+    // create a 5 feature dataset and train the model
+    cv::Ptr<RTrees> model = RTrees::create();
+    Mat samples(10, numFeatures, CV_32F);
+    randu(samples, 0, 10);
+    Mat labels = (Mat_<int>(10,1) << 0,0,0,0,0,1,1,1,1,1);
+    cv::Ptr<TrainData> trainData = TrainData::create(samples, cv::ml::ROW_SAMPLE, labels);
+    model->train(trainData);
+    // try to predict on data which have fewer features - this should throw an exception
+    for(int i = 1; i < numFeatures - 1; ++i) {
+        Mat test(1, i, CV_32FC1);
+        ASSERT_THROW(model->predict(test), Exception);
+    }
+    // try to predict on data which have more features - this should also throw an exception
+    Mat test(1, numFeatures + 1, CV_32FC1);
+    ASSERT_THROW(model->predict(test), Exception);
+}
 
 
 }} // namespace