Yunet Model gaves an unexpected output sometimes

C++
1

In the context of creating an application to extract Face Rectangles coords of a great number of faces, in a file path, in the same size dimensions, I got an issue with Yunet.
Code example:

for( ... ) // iterates in all images of the filepat
...
FaceDetector->setInputSize(image.size());
cv::Mat faces(1, 15, CV_32FC1, cv::Scalar(0));
FaceDetector->detect(image, faces);
...

In docs of Yunet usage, it says the expected output for this model is a [num_faces, 15]. But in some random cases, I have only one face and the model gaves me [2 x 15] with infinitely large numbers at the first 15 positions and the expected 15 numbers after these.

Unexpected output for example: [-inf, -inf, inf, inf, -7.3996823e+29, 8.3138902e+29, -6.9410597e+29, 9.5680955e+29, -3.6332516e+29, 6.361716e+29, -5.7762872e+29, 8.271784e+29, -4.8886777e+29, 9.2385431e+29, 1;
8.0682755, 1.9483109, 89.627586, 123.62176, 34.569054, 43.413845, 74.630112, 48.56892, 48.431858, 69.483932, 30.405964, 89.4823, 65.017296, 93.899574, 0.99998468]

In someway, it looks kinda the model find a true face in the infinity (a position completely out of my image).
Note: all my images have the same dimension, an image contains exactly one person and the face it’s totally in the expected range of Yunet input’s " faces of pixels between around 10x10 to 300x30". This bug occurs for different images in different executions. Example: in my first execution, the bug appeared for the image10, the next execution my image10 gaves me the correct [1x15] and then the bug appeared in image94. Maybe it could be a garbage initialization issue, but I always initialize my cv::Mat faces with zero, so I don’t know if the reason could still be related to this way… Any idea?

2

which exact opencv version and yunet model ?

unlikely the cause of your problem, but please don’t pre-allocate this:

it’s going to be overwritten anyway and you’re fooling yourself believing something about type, shape or content like this

3

My model version was deprecated. Thanks for answering. It seems the new model version fixed it.

4

I’m seeing the same thing with the face_detection_yunet_2023mar.onnx model from the model zoo… Any pointer on what it could be? Is there a newer model than this one available?