Feature Extraction CUDA/PYTHON

Python
1

Hello, I’m quite new to cuda, I need to find the characteristic points between two successive images and I don’t know how to do it, I have 4 images and I need to extract the characteristic points of the image 0.png in 1.png, search for those points features of 1.png in 2.png and so on (a visual odometry algorithm), and I get the following error, can someone help me


import cv2 as cv

img_files = ['0.png', '1.png', '2.png', '3.png']
orb = cv.cuda.ORB_create(5000)
matcherGPU = cv.cuda.DescriptorMatcher_createBFMatcher(cv.NORM_HAMMING) 

gpu_frame = cv.cuda_GpuMat()

for i in range(len(img_files)):
    screenshot = cv.imread(f"C:\\Users\\Geddy\\Desktop\\{img_files[i]}")
    gpu_frame.upload(screenshot)  
    screenshot = cv.cuda.cvtColor(gpu_frame, cv.COLOR_RGB2BGR)
    
    kp1, des1 = orb.detectAndComputeAsync(screenshot[i-1], None)
    kp2, des2 = orb.detectAndComputeAsync(screenshot[i], None)
    
    matches = matcherGPU.knnMatch(des1,des2, k=2)
    # download image from GPU (cv2.cuda_GpuMat -> numpy.ndarray)
    screenshot = screenshot.download()

…

Traceback (most recent call last):
  File "c:\Users\Geddy\Desktop\test2.py", line 20, in <module>
    kp1, des1 = orb.detectAndComputeAsync(screenshot[i-1], None) # Both are returned as GPU Mats
TypeError: 'cv2.cuda.GpuMat' object is not subscriptable
2

A GpuMat is an OpenCV object which can only be accessed using OpenCV api functions, your confusing it with a numpy array, pytorch tensor etc. If you want to access a single color plane then you will need to split it first

or if you just want to process the grey level frame cvtColor to grey

3

Hello, thanks for your reply, what I want is to extract the characteristics of different images and do the matching (it is a visual odometry algorithm), I want to know how to do this algorithm with different images, not only with 2 (it is easy with 2) , the problem is when I want to do it with different images, suppose I have 10 images (a sequence of the KITTI dataset), I need to search the characteristics of image 1, with 2, then 2 with 3, 3 with 4 and so on, below I put an example with 2 images that works, but I don’t know how to do it for several images

from distutils.command.upload import upload
import cv2
import numpy as np

        
MAX_FEATURES = 500
GOOD_MATCH_PERCENT = 0.15
        # load images into numpy
npMat1 = cv2.imread("C:\\Users\\Geddy\\OneDrive\\Desktop\\0.png")
npMat2 = cv2.imread("C:\\Users\\Geddy\\OneDrive\\Desktop\\1.png")      
        # upload into Cuda
cuMat1 = cv2.cuda_GpuMat()
cuMat2 = cv2.cuda_GpuMat()
cuMat1.upload(npMat1)
cuMat2.upload(npMat2)        
        #convert to Gray
cuMat1g = cv2.cuda.cvtColor(cuMat1, cv2.COLOR_RGB2GRAY)
cuMat2g = cv2.cuda.cvtColor(cuMat2, cv2.COLOR_RGB2GRAY)
        #ORB
corb = cv2.cuda_ORB.create(MAX_FEATURES)
_kps1, _descs1 = corb.detectAndComputeAsync(cuMat1g, None)
_kps2, _descs2 = corb.detectAndComputeAsync(cuMat2g, None)
        #convert Keypoints to CPU
kps1 = [cv2.KeyPoint() for i in range(MAX_FEATURES)]
kps2 = [cv2.KeyPoint() for i in range(MAX_FEATURES)]
corb.convert(_kps1)
corb.convert(_kps2)      
        # Matching
cbf = cv2.cuda_DescriptorMatcher.createBFMatcher(cv2.NORM_HAMMING)
cmatches = cbf.match(_descs1, _descs2) 
        # Sort matches by score
cmatches=sorted(cmatches, key=lambda x: x.distance, reverse=False)
        # Remove not so good matches
numGoodMatches = int(len(cmatches) * GOOD_MATCH_PERCENT)
cmatches = cmatches[:numGoodMatches]  
        # Draw top matches
imMatches = cv2.drawMatches(npMat1, kps1, npMat2, kps2, cmatches, None)
cv2.imwrite("gpu_matches.jpg", imMatches)
4

It is easy to create a GPU_MATH type variable, because it would be 2 images, but when there are 200 images, what do I do? could you call the upload function and upload 200 images? or do I have to upload image by image (as shown in the example)?

5

You can do either but I would recommend uploading the images just before you use them to avoid using a lot of GPU memory.

6

Thank you very much, even so I have no idea how to do it, I would have an array type variable that would contain the 10 images, I would create a variable cuMat1 = cv2.cuda_GpuMat() and then I would have to load those 10 images, I have no idea there How to do it, I don’t know if I have to create a cuMat1 = cv2.cuda_GpuMat() for each of the images (as in the previous example, 2 are created because there are 2 images). Example:

npMat1 = cv2.imread("C:\\Users\\Geddy\\OneDrive\\Desktop\\*.png") #list of 10 images
cuMat1 = cv2.cuda_GpuMat()
cuMat1.upload(npMat1)

This is not working, and if I put everything inside a for it doesn’t work either, so it generates the following question: is it possible to create a single variable cuMat1 = cv2.cuda_GpuMat() and then when the function cuMat1.upload(npMat1 ), load an array (because the variable contains 10 images)? Or would it be necessary to create a cuMat1 = cv2.cuda_GpuMat() for each of the images?

7

You need to get a list of all the images then iterate over it uploading each one individually. I think this is more of a python question than an OpenCV one.

8

imread() only reads a single image, not a list of images.

if you want to do that, use e.g. glob.glob(directory) to get a list of filenames, then imread() those one by one.

9

Sorry for my mistake, I know that imread is only for an image I just wanted to exemplify what I would like to do, I will try to upload image by image to see if it works