Merge pull request #25810 from fengyuentau:python/fix_parsing_3d_mat_in_dnn

python: attempts to fix 3d mat parsing problem for dnn #25810

Fixes https://github.com/opencv/opencv/issues/25762 https://github.com/opencv/opencv/issues/23242
Relates https://github.com/opencv/opencv/issues/25763 https://github.com/opencv/opencv/issues/19091

Although `cv.Mat` has already been introduced to workaround this problem, people do not know it and it kind of leads to confusion with `numpy.array`. This patch adds a "switch" to turn off the auto multichannel feature when the API is from cv::dnn::Net (more specifically, `setInput`) and the parameter is of type `Mat`. This patch only leads to changes of three places in `pyopencv_generated_types_content.h`:

```.diff
static PyObject* pyopencv_cv_dnn_dnn_Net_setInput(PyObject* self, PyObject* py_args, PyObject* kw)
{
...
- pyopencv_to_safe(pyobj_blob, blob, ArgInfo("blob", 0)) &&
+ pyopencv_to_safe(pyobj_blob, blob, ArgInfo("blob", 8)) &&
...
}

// I guess we also need to change this as one-channel blob is expected for param
static PyObject* pyopencv_cv_dnn_dnn_Net_setParam(PyObject* self, PyObject* py_args, PyObject* kw)
{
...
- pyopencv_to_safe(pyobj_blob, blob, ArgInfo("blob", 0)) )
+ pyopencv_to_safe(pyobj_blob, blob, ArgInfo("blob", 8)) )
...
- pyopencv_to_safe(pyobj_blob, blob, ArgInfo("blob", 0)) )
+ pyopencv_to_safe(pyobj_blob, blob, ArgInfo("blob", 8)) )
...
}
```

Others are unchanged, e.g. `dnn_SegmentationModel` and stuff like that.

### Pull Request Readiness Checklist

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
      Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake

modules/core/include/opencv2/core/cvdef.h

@@ -478,6 +478,7 @@ Cv64suf;
 #define CV_OUT
 #define CV_PROP
 #define CV_PROP_RW
+#define CV_ND // Indicates that input data should be parsed into Mat without channels
 #define CV_WRAP
 #define CV_WRAP_AS(synonym)
 #define CV_WRAP_MAPPABLE(mappable)

modules/dnn/include/opencv2/dnn/dnn.hpp

@@ -642,13 +642,13 @@ CV__DNN_INLINE_NS_BEGIN
          *  @param outputName name for layer which output is needed to get
          *  @details If @p outputName is empty, runs forward pass for the whole network.
          */
-        CV_WRAP void forward(OutputArrayOfArrays outputBlobs, const String& outputName = String());
+        CV_WRAP void forward(CV_ND OutputArrayOfArrays outputBlobs, const String& outputName = String());
 
         /** @brief Runs forward pass to compute outputs of layers listed in @p outBlobNames.
          *  @param outputBlobs contains blobs for first outputs of specified layers.
          *  @param outBlobNames names for layers which outputs are needed to get
          */
-        CV_WRAP void forward(OutputArrayOfArrays outputBlobs,
+        CV_WRAP void forward(CV_ND OutputArrayOfArrays outputBlobs,
                              const std::vector<String>& outBlobNames);
 
         /** @brief Runs forward pass to compute outputs of layers listed in @p outBlobNames.
@@ -727,7 +727,7 @@ CV__DNN_INLINE_NS_BEGIN
          *  as:
          * \f[input(n,c,h,w) = scalefactor \times (blob(n,c,h,w) - mean_c)\f]
          */
-        CV_WRAP void setInput(InputArray blob, const String& name = "",
+        CV_WRAP void setInput(CV_ND InputArray blob, const String& name = "",
                               double scalefactor = 1.0, const Scalar& mean = Scalar());
 
         /** @brief Sets the new value for the learned param of the layer.
@@ -738,8 +738,8 @@ CV__DNN_INLINE_NS_BEGIN
          *  @note If shape of the new blob differs from the previous shape,
          *  then the following forward pass may fail.
         */
-        CV_WRAP void setParam(int layer, int numParam, const Mat &blob);
-        CV_WRAP inline void setParam(const String& layerName, int numParam, const Mat &blob) { return setParam(getLayerId(layerName), numParam, blob); }
+        CV_WRAP void setParam(int layer, int numParam, CV_ND const Mat &blob);
+        CV_WRAP inline void setParam(const String& layerName, int numParam, CV_ND const Mat &blob) { return setParam(getLayerId(layerName), numParam, blob); }
 
         /** @brief Returns parameter blob of the layer.
          *  @param layer name or id of the layer.

modules/dnn/misc/python/test/test_dnn.py

@@ -455,10 +455,6 @@ class dnn_test(NewOpenCVTests):
                                     "Verify OPENCV_DNN_TEST_DATA_PATH configuration parameter.")
 
         input = np.load(input_file)
-        # we have to expand the shape of input tensor because Python bindings cut 3D tensors to 2D
-        # it should be fixed in future. see : https://github.com/opencv/opencv/issues/19091
-        # please remove `expand_dims` after that
-        input = np.expand_dims(input, axis=3)
         gold_output = np.load(output_file)
 
         for backend, target in self.dnnBackendsAndTargets:
@@ -469,10 +465,63 @@ class dnn_test(NewOpenCVTests):
             net.setPreferableBackend(backend)
             net.setPreferableTarget(target)
 
+            # Check whether 3d shape is parsed correctly for setInput
             net.setInput(input)
-            real_output = net.forward()
 
-            normAssert(self, real_output, gold_output, "", getDefaultThreshold(target))
+            # Case 0: test API `forward(const String& outputName = String()`
+            real_output = net.forward() # Retval is a np.array of shape [2, 5, 3]
+            normAssert(self, real_output, gold_output, "Case 1", getDefaultThreshold(target))
+
+            '''
+            Pre-allocate output memory with correct shape.
+            Normally Python users do not use in this way,
+            but we have to test it since we design API in this way
+            '''
+            # Case 1: a np.array with a string of output name.
+            #         It tests API `forward(OutputArrayOfArrays outputBlobs, const String& outputName = String()`
+            #         when outputBlobs is a np.array and we expect it to be the only output.
+            real_output = np.empty([2, 5, 3], dtype=np.float32)
+            real_output = net.forward(real_output, "237") # Retval is a tuple with a np.array of shape [2, 5, 3]
+            normAssert(self, real_output, gold_output, "Case 1", getDefaultThreshold(target))
+
+            # Case 2: a tuple of np.array with a string of output name.
+            #         It tests API `forward(OutputArrayOfArrays outputBlobs, const String& outputName = String()`
+            #         when outputBlobs is a container of several np.array and we expect to save all outputs accordingly.
+            real_output = tuple(np.empty([2, 5, 3], dtype=np.float32))
+            real_output = net.forward(real_output, "237") # Retval is a tuple with a np.array of shape [2, 5, 3]
+            normAssert(self, real_output, gold_output, "Case 2", getDefaultThreshold(target))
+
+            # Case 3: a tuple of np.array with a string of output name.
+            #         It tests API `forward(OutputArrayOfArrays outputBlobs, const std::vector<String>& outBlobNames)`
+            real_output = tuple(np.empty([2, 5, 3], dtype=np.float32))
+            # Note that it does not support parsing a list , e.g. ["237"]
+            real_output = net.forward(real_output, ("237")) # Retval is a tuple with a np.array of shape [2, 5, 3]
+            normAssert(self, real_output, gold_output, "Case 3", getDefaultThreshold(target))
+
+    def test_set_param_3d(self):
+        model_path = self.find_dnn_file('dnn/onnx/models/matmul_3d_init.onnx')
+        input_file = self.find_dnn_file('dnn/onnx/data/input_matmul_3d_init.npy')
+        output_file = self.find_dnn_file('dnn/onnx/data/output_matmul_3d_init.npy')
+
+        input = np.load(input_file)
+        output = np.load(output_file)
+
+        for backend, target in self.dnnBackendsAndTargets:
+            printParams(backend, target)
+
+            net = cv.dnn.readNet(model_path)
+
+            node_name = net.getLayerNames()[0]
+            w = net.getParam(node_name, 0) # returns the original tensor of three-dimensional shape
+            net.setParam(node_name, 0, w)  # set param once again to see whether tensor is converted with correct shape
+
+            net.setPreferableBackend(backend)
+            net.setPreferableTarget(target)
+
+            net.setInput(input)
+            res_output = net.forward()
+
+            normAssert(self, output, res_output, "", getDefaultThreshold(target))
 
     def test_scalefactor_assign(self):
         params = cv.dnn.Image2BlobParams()

modules/python/src2/cv2.hpp

@@ -46,19 +46,22 @@ private:
     static const uint32_t arg_outputarg_flag     = 0x1;
     static const uint32_t arg_arithm_op_src_flag = 0x2;
     static const uint32_t arg_pathlike_flag      = 0x4;
+    static const uint32_t arg_nd_mat_flag        = 0x8;
 
 public:
     const char* name;
     bool outputarg;
     bool arithm_op_src;
     bool pathlike;
+    bool nd_mat;
     // more fields may be added if necessary
 
     ArgInfo(const char* name_, uint32_t arg_) :
         name(name_),
         outputarg((arg_ & arg_outputarg_flag) != 0),
         arithm_op_src((arg_ & arg_arithm_op_src_flag) != 0),
-        pathlike((arg_ & arg_pathlike_flag) != 0) {}
+        pathlike((arg_ & arg_pathlike_flag) != 0),
+        nd_mat((arg_ & arg_nd_mat_flag) != 0) {}
 
 private:
     ArgInfo(const ArgInfo&) = delete;

modules/python/src2/cv2_convert.cpp

@@ -173,7 +173,7 @@ bool pyopencv_to(PyObject* o, Mat& m, const ArgInfo& info)
 
     CV_LOG_DEBUG(NULL, "Incoming ndarray '" << info.name << "': ndims=" << ndims << "  _sizes=" << pycv_dumpArray(_sizes, ndims) << "  _strides=" << pycv_dumpArray(_strides, ndims));
 
-    bool ismultichannel = ndims == 3 && _sizes[2] <= CV_CN_MAX;
+    bool ismultichannel = ndims == 3 && _sizes[2] <= CV_CN_MAX && !info.nd_mat;
     if (pyopencv_Mat_TypePtr && PyObject_TypeCheck(o, pyopencv_Mat_TypePtr))
     {
         bool wrapChannels = false;

modules/python/src2/cv2_convert.hpp

@@ -340,22 +340,38 @@ static bool pyopencv_to_generic_vec(PyObject* obj, std::vector<Tp>& value, const
     {
         return true;
     }
-    if (!PySequence_Check(obj))
+    if (info.nd_mat && PyArray_Check(obj))
     {
-        failmsg("Can't parse '%s'. Input argument doesn't provide sequence protocol", info.name);
-        return false;
-    }
-    const size_t n = static_cast<size_t>(PySequence_Size(obj));
-    value.resize(n);
-    for (size_t i = 0; i < n; i++)
-    {
-        SafeSeqItem item_wrap(obj, i);
-        if (!pyopencv_to(item_wrap.item, value[i], info))
+        /*
+            If obj is marked as nd mat and of array type, it is parsed to a single
+            mat in the target vector to avoid being split into multiple mats
+        */
+        value.resize(1);
+        if (!pyopencv_to(obj, value.front(), info))
         {
-            failmsg("Can't parse '%s'. Sequence item with index %lu has a wrong type", info.name, i);
+            failmsg("Can't parse '%s'. Array item has a wrong type", info.name);
             return false;
         }
     }
+    else // parse as sequence
+    {
+        if (!PySequence_Check(obj))
+        {
+            failmsg("Can't parse '%s'. Input argument doesn't provide sequence protocol", info.name);
+            return false;
+        }
+        const size_t n = static_cast<size_t>(PySequence_Size(obj));
+        value.resize(n);
+        for (size_t i = 0; i < n; i++)
+        {
+            SafeSeqItem item_wrap(obj, i);
+            if (!pyopencv_to(item_wrap.item, value[i], info))
+            {
+                failmsg("Can't parse '%s'. Sequence item with index %lu has a wrong type", info.name, i);
+                return false;
+            }
+        }
+    }
     return true;
 }
 

modules/python/src2/gen2.py

@@ -488,6 +488,10 @@ class ArgInfo(object):
             return self.name + '_'
         return self.name
 
+    @property
+    def nd_mat(self):
+        return '/ND' in self._modifiers
+
     @property
     def inputarg(self):
         return '/O' not in self._modifiers
@@ -528,6 +532,7 @@ class ArgInfo(object):
         arg  = 0x01 if self.outputarg else 0x0
         arg += 0x02 if self.arithm_op_src_arg else 0x0
         arg += 0x04 if self.pathlike else 0x0
+        arg += 0x08 if self.nd_mat else 0x0
         return "ArgInfo(\"%s\", %d)" % (self.name, arg)
 
 

modules/python/src2/hdr_parser.py

@@ -82,6 +82,10 @@ class CppHeaderParser(object):
         modlist = []
 
         # pass 0: extracts the modifiers
+        if "CV_ND" in arg_str:
+            modlist.append("/ND")
+            arg_str = arg_str.replace("CV_ND", "")
+
         if "CV_OUT" in arg_str:
             modlist.append("/O")
             arg_str = arg_str.replace("CV_OUT", "")