I am unable to run the command detail_ImageFeatures().descriptors in Python

[VisheshKhosla]

System Version(Information)

• OpenCV =>4.1.2 (Python 3.7.3)
• Operating System / Platform => Windows 10 64 Bit
• Hardware => Dell Inspiron 7460
• Compiler/Interpreter => Jupyter Notebooks
• GPU =>NVIDIA 940MX

Detailed Description

I am unable to access a public data member of cv::detail::ImageFeatures::descriptors. When i run this command:
import cv2 obj_Umat=cv2.detail_ImageFeatures().descriptors

The command cv2.detail_ImageFeatures().descriptors returns a UMat object.
My interpreter then gives me a response that :

The kernel appears to have died

[alalek]

Please clarify what are you trying to do? What is expected result?

[VisheshKhosla]

I am not aiming to do something,but why is my code not working? I was just trying out these data members present in this documentation: https://docs.opencv.org/3.4/d4/db5/structcv_1_1detail_1_1ImageFeatures.html#details
All are working fine except descriptors and keypoints.

[czgdp1807]

Hi,
I tried the code above. It looks like for me, bindings for cv::detail::ImageFeatures. I checked build-opencv/modules/python_bindings_generator/pyopencv_signatures.json and observed that there is not key for cv::detail::ImageFeatures. Is that the underlying issue? What was the error when you tried to execute the above line. Thanks.

[alalek]

• dir(image_features_object) ?
• image_features_object.getKeypoints() ?

[saskatchewancatch]

@czgdp1807 Not in the Python2 lib. In Python3, yes. You will see the key when you import cv2 in Python3 interpreter. For OP, it seems the underlying issue is that accessing the descriptors attribute of the struct results in a seg fault.

@alalek

Python 3.6.9 (default, Nov  7 2019, 10:44:02) 
[GCC 8.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import cv2
>>> f = cv2.detail_ImageFeatures()
>>> dir(f)
['__class__', '__delattr__', '__dir__', '__doc__', '__eq__', '__format__', '__ge__', '__getattribute__', '__gt__', '__hash__', '__init__', '__init_subclass__', '__le__', '__lt__', '__ne__', '__new__', '__reduce__', '__reduce_ex__', '__repr__', '__setattr__', '__sizeof__', '__str__', '__subclasshook__', 'descriptors', 'getKeypoints', 'img_idx', 'img_size']
>>> f.getKeypoints
<built-in method getKeypoints of cv2.detail_ImageFeatures object at 0x7fce3dc3fbe0>
>>> f.getKeypoints()
[]
>>> f.descriptors()
Segmentation fault (core dumped)

I don't use Stitching module, and I don't use the Python API, but it seems to me I should get an empty numpy array or None type instead of segfault.

I'm running a Ubuntu 18.04 on a VM with no GPU. Still, an unset UMat attribute should return None even if no reliance on OpenCL. I wonder if all UMat's members in structs accessed through Python3 api cause such segfaults.

[rllyryan]

Any leads on this? I am trying to transfer information from features detected by SIFT into a detail_ImageFeatures object.

[Sanyam-Mehta]

@czgdp1807 Not in the Python2 lib. In Python3, yes. You will see the key when you import cv2 in Python3 interpreter. For OP, it seems the underlying issue is that accessing the descriptors attribute of the struct results in a seg fault.

@alalek

Python 3.6.9 (default, Nov  7 2019, 10:44:02) 
[GCC 8.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import cv2
>>> f = cv2.detail_ImageFeatures()
>>> dir(f)
['__class__', '__delattr__', '__dir__', '__doc__', '__eq__', '__format__', '__ge__', '__getattribute__', '__gt__', '__hash__', '__init__', '__init_subclass__', '__le__', '__lt__', '__ne__', '__new__', '__reduce__', '__reduce_ex__', '__repr__', '__setattr__', '__sizeof__', '__str__', '__subclasshook__', 'descriptors', 'getKeypoints', 'img_idx', 'img_size']
>>> f.getKeypoints
<built-in method getKeypoints of cv2.detail_ImageFeatures object at 0x7fce3dc3fbe0>
>>> f.getKeypoints()
[]
>>> f.descriptors()
Segmentation fault (core dumped)

I don't use Stitching module, and I don't use the Python API, but it seems to me I should get an empty numpy array or None type instead of segfault.

I'm running a Ubuntu 18.04 on a VM with no GPU. Still, an unset UMat attribute should return None even if no reliance on OpenCL. I wonder if all UMat's members in structs accessed through Python3 api cause such segfaults.

I am also seeing a segfault when I try to access the descriptors. Are there any updates on this?

[berak]

related: #20979

hmm it generates code like this:

static PyObject* pyopencv_detail_ImageFeatures_get_descriptors(pyopencv_detail_ImageFeatures_t* p, void *closure)
{
    return pyopencv_from(p->v.descriptors);
}

but i cant find pyopencv_from(const UMat &m) anywhere.

[xmba15]

This is a little hack to create an empty cv2.detail_ImageFeatures and access its descriptors

image_features = cv2.detail.computeImageFeatures2(
      cv2.SIFT_create(),
      np.zeros_like((0, 0), dtype=np.uint8),
      np.zeros_like((0, 0), dtype=np.uint8),
)
image_features.descriptors # of type cv2.UMat

# update descriptors:
# image_features.descriptors = cv2.UMat(some numpy array)

After that, you can also update image_features ' keypoints, img_size