opencv facerecognition: subspaceproject and subspa

问题:

I have searched through the web, but haven't found any answer so far to the following question, therefore I want to ask here if someone could help me out with that:

basically what i need is the same as in the solution from PraveenofPersia/Jesse there, but only the python implementation considering a Fisherface Recognizer: Any tips on confidence score for face verification (as opposed to face recognition)?

up to now I am facing the problem, that cv2 does not offer either subspaceProject nor any other.

has anyone suggestions here?

thank you!

回答1:

those functions are unfortunately not exposed to the python api by default.

if you're building the cv2.pyd from source, there's an easy remedy :

• locate their resp. declarations, opencv/modules/contrib/include/opencv2/contrib/contrib.hpp
• change CV_EXPORTS Mat subspaceProject(...) to CV_EXPORTS_W Mat subspaceProject(...)

• change CV_EXPORTS Mat subspaceReconstruct(...) to CV_EXPORTS_W Mat subspaceReconstruct(...)

• rerun cmake / make to rebuild the cv libs and the python wrapper module

thew additional _W prefix will add those functions to the generated wrappers

回答2:

I went ahead and re-wrote the functions in C++ as python. It's not the cleanest, but it works! If you take this python code and couple it with the high level concepts from another C++ example you can do exactly what you're trying to do.

 # projects samples into the LDA subspace
 def subspace_project(eigenvectors_column, mean, source):
     source_rows = len(source)
     source_cols = len(source[0])

     if len(eigenvectors_column) != source_cols * source_rows:
         raise Exception("wrong shape")

     flattened_source = []
     for row in source:
         flattened_source += [float(num) for num in row]
     flattened_source = np.asarray(flattened_source)
     delta_from_mean = cv2.subtract(flattened_source, mean)
     # flatten the matrix then convert to 1 row by many columns
     delta_from_mean = np.asarray([np.hstack(delta_from_mean)])

     empty_mat = np.array(eigenvectors_column, copy=True)  # this is required for the function call but unused
     result = cv2.gemm(delta_from_mean, eigenvectors_column, 1.0, empty_mat, 0.0)
     return result


 # reconstructs projections from the LDA subspace
 def subspace_reconstruct(eigenvectors_column, mean, projection, image_width, image_height):
     if len(eigenvectors_column[0]) != len(projection[0]):
         raise Exception("wrong shape")

     empty_mat = np.array(eigenvectors_column, copy=True)  # this is required for the function call but unused
     # GEMM_2_T transposes the eigenvector
     result = cv2.gemm(projection, eigenvectors_column, 1.0, empty_mat, 0.0, flags=cv2.GEMM_2_T)

     flattened_array = result[0]
     flattened_image = np.hstack(cv2.add(flattened_array, mean))
     flattened_image = np.asarray([np.uint8(num) for num in flattened_image])
     all_rows = []
     for row_index in xrange(image_height):
         row = flattened_image[row_index * image_width: (row_index + 1) * image_width]
         all_rows.append(row)
     image_matrix = np.asarray(all_rows)
     image = normalize_hist(image_matrix)
     return image


 def normalize_hist(face):
     face_as_mat = np.asarray(face)
     equalized_face = cv2.equalizeHist(face_as_mat)
     equalized_face = cv.fromarray(equalized_face)                                                                                                                                                                                 
     return equalized_face