How to get hog features at a keypoint using python?

Python
, ,
1

Hi,
I’m trying to get the hog features at specific keypoints of an image.
I’m resizing the images to size 256x256x3
then, I initialize
hog = cv2.HOGDescriptor(winSize=(256, 256), blockSize=(16,16), blockStride=(8,8), cellSize=(8,8), nbins=9)
hog.compute(image) → this returns a vector of length 34596

if I’m using ‘locations’

hog.compute(image, locations=[points]) it returns a vector of length 34596xnum_points. but for some reason this vector is all 0’s.

What am I doing wrong?

appreciate any help

2

Hi,
Can you give a full minimal example?

4

Screenshot from 2021-02-16 16-37-10
Screenshot from 2021-02-16 16-37-10898×774 94.6 KB

Hope it’s clearer

5

I’m a little lazy. Can you post your code as text (it’s better for google too) and use an opencv image

7

using lena

import numpy as np
import cv2
import matplotlib.pyplot as plt

lena = cv2.imread('lena.jpg')
lena = cv2.resize(lena, (256, 256)) # BGR image normalized to size of 256x256, no other preprocessing
points = [(155, 160), (135, 175), (160, 175), (165, 135), (135, 135)] # 5 keypoints of intrest

lena_p = lena # just for drawing points
for (x, y) in points:
    cv2.circle(lena_p, (x, y), 2, (255, 0, 0), 2)	

plt.figure(figsize=(7, 7))
plt.imshow(lena_p[..., ::-1])
plt.xticks([]), plt.yticks([])  # to hide tick values on X and Y axis
plt.show()

# HOG params
winSize = (256, 256)
blockSize = (16,16)
blockStride = (8,8)
cellSize = (8,8)
nbins = 9

# initialize hog descriptor
hog = cv2.HOGDescriptor(winSize, blockSize, blockStride, cellSize, nbins)
# compute hog features
fv = hog.compute(lena)  # computing the feature vector fv
print(fv, np.all(fv==0)) # the fv values
print(f'fv dimensions {fv.shape}, {fv.shape[0]} == 31*31*2*2*9 :', fv.shape[0] == 31*31*2*2*9)
print()

fv_lm = hog.compute(lena, locations=points)  # computing the feature vector for the 5 points
print(fv_lm, np.all(fv_lm==0)) # the fv values
print(f'fv_lm dimensions {fv_lm.shape}, {fv_lm.shape[0]} / {len(points)} == {fv.shape[0]} :', 
      fv.shape[0] == fv_lm.shape[0] / len(points))
8

hint: cv.samples.findFile("lena.jpg") is a neat utility function that gives you the path to a resource file that comes with OpenCV.

1 Like
9

I must say that I don’t know how to use hog.
I run same program in C++.

Mat img = imread("g:/lib/opencv/samples/data/lena.jpg", IMREAD_GRAYSCALE);
Mat lena;
resize(img, lena, Size(256, 256));
Size winSize(256, 256);
Size blockSize(16, 16);
Size blockStride(8, 8);
Size cellSize(8, 8);
int nbins = 9;
vector <Point> pt = { Point(10,10), Point(100,20), Point(150,30), Point(150,30) };
vector< float> descriptors;
HOGDescriptor hog(winSize, blockSize, blockStride, cellSize, nbins);

hog.compute(lena, descriptors);
hog.compute(lena, descriptors,Size(), Size(), pt);

call hog.compute(lena, descriptors); 34596 desciptors for [0] is 0.226208031, [10] is 0.163035244 [34595] is 0.0669640526
hog.compute(lena, descriptors,Size(), Size(), pt); 138384 descriptors [0], [10] and [34595] are same and after all values are 0

I think that my program is wrong example…

10

I dig into the code and this condition must be false to get non zero descriptors

if( pt0.x < -padding.width || pt0.x > img.cols + padding.width - winSize.width ||
    pt0.y < -padding.height || pt0.y > img.rows + padding.height - winSize.height )

Hence you must set padding and winSize in compute args

11

I appreciate the help, thanks.
My mistake was setting winSize = (256, 256). hence for each keypoint it tried to calculate 31x31x2x2x9 histograms (and returned 0’s because the stride and padding issue).
Once I set winSize = (16, 16), for each keypoint it calculated the correct number of histograms 1x1x2x2x9 in the surrounding of the keypoint.

Thanks