OpenCV(4.8.0) Error: The function/feature is not implemented () in cv::debug_build_guard::_OutputArray::assign

I'm trying to translate some Python code to C++, and I run into this code:

self.net = cv.dnn.readNetFromDarknet(cfg_file, weights_file)
self.net.setPreferableBackend(cv.dnn.DNN_BACKEND_OPENCV)
self.ln = self.net.getLayerNames()
self.ln = [self.ln[i-1] for i in self.net.getUnconnectedOutLayers()]
     ...
blob = cv.dnn.blobFromImage(img, 1/255.0, (416, 416), swapRB=True, crop=False)
self.net.setInput(blob)
outputs = self.net.forward(self.ln)
outputs = np.vstack(outputs)

As you can see, the code is taking the output from net forward from the unconnected layer names that are specified in self.ln. Now here is my C++ code:

static Mat blobFromImg;
bool swapRB = true;
blobFromImage(img, blobFromImg, 1/255.0, Size(416, 416), Scalar(), swapRB, false);
network.setInput(blobFromImg);
Mat outMat;
network.forward(outMat , network.getUnconnectedOutLayersNames());
vconcat(outMat, outMat);

As you can see, I am using cv::vconcat in place of np.vstack instead of making a new std::vector<<cv::String>> to pass to network.forward(). I use getUnconnectedOutLayersNames(). Yet I get this error:

OpenCV(4.8.0) Error: The function/feature is not implemented () in cv::debug_build_guard::_OutputArray::assign, file D:\vcpkg\buildtrees\opencv4\src\4.8.0-2bf495557d.clean\modules\core\src\matrix_wrap.cpp, line 2052

The line where the exception occurs is network.forward(). It is because of this method that network.getUnconnectedOutLayersNames() doesn't give the exception by itself.

If I do:

std::vector<cv::String> ln;
auto layers = network.getLayerNames();
for (auto i : network.getUnconnectedOutLayers()){
   ln.push_back(layers[i-1]);
}
   ...
network.setInput(blobFromImg);
Mat outMat;
network.forward(outMat , ln);
vconcat(outMat, outMat);

I get the same result.

Why is the function not implemented? Why does Python have this implementation.

Note: I am using darknet/yolo.

Solution

Based on the context, we're dealing with the following overload of cv::dnn::Net::forward, the documentation of which says:

void cv::dnn::Net::forward(OutputArrayOfArrays outputBlobs, const std::vector<String>& outBlobNames)

Parameters:

outputBlobs contains blobs for first outputs of specified layers.

outBlobNames names for layers which outputs are needed to get

The problem is with the first parameter outputBlobs. The documentation tells us about cv::OutputArrayOfArrays that it's an alias for cv::_OutputArray :

typedef OutputArray OutputArrayOfArrays;

Other than that, there's not much to be found on that page. For the more useful information, we need to look at the documentation of cv::_InputArray that it derives from:

... can be constructed from ..., std::vector<std::vector<T>>, std::vector<Mat>, std::vector<Mat_<T>>

...

There is another related type, InputArrayOfArrays, which is currently defined as a synonym for InputArray:

typedef InputArray InputArrayOfArrays;

It denotes function arguments that are either vectors of vectors or vectors of matrices. A separate synonym is needed to generate Python/Java etc. wrappers properly. At the function implementation level their use is similar, but _InputArray::getMat(idx) should be used to get header for the idx-th component of the outer vector and _InputArray::size().area() should be used to find the number of components (vectors/matrices) of the outer vector.

And suddenly it becomes clear. We're calling the function with a sequence a layer names, we should be getting back a sequence of blobs. We know that a blob can be represented as a 4D cv::Mat. A sequence of those in C++ would be a vector of Mats: std::vector<cv::Mat>. So, change the first parameter of forward to be that type. And as you say "that worked perfectly".