An ML model that classifies yoga pose into 4 most famous asanas namely downward dog, plank pose, tree pose, goddess pose, and warrior-2 pose using Mediapipe Blazepose for feature extraction.


Dataset is a combined Dataset of :

Preprocessing Images

Rescaling images

Images are first resized to reduce computation.

def resize_2D_array(overall_images):
new_arr_for_outerdirectory = []
for i in range(len(overall_images)):
col = dict[i]
new_array_for_subdirectory = []
for j in range(len(overall_images[i])):
img = overall_images[i][j]
h, w = img.shape[:2]
if h < w:
img = cv2.resize(img, (DESIRED_WIDTH, math.floor(h/(w/DESIRED_WIDTH))))
img = cv2.resize(img, (math.floor(w/(h/DESIRED_HEIGHT)), DESIRED_HEIGHT))
#appending the image to a new array
return new_arr_for_outerdirectory

Brightness Adjustment

Gamma correction is a non-linear adjustment to individual pixel values. In image normalization, linear operations are carried out on individual pixels, gamma correction carries out a non-linear operation on the source image pixels, and can cause saturation of the image to bealtered.

#for brightness improvementdef gammaCorrection(src, gamma):
invGamma = 1 / gamma

table = [((i / 255) ** invGamma) * 255 for i in range(256)]
table = np.array(table, np.uint8)

return cv2.LUT(src, table)
def isbright(image, dim=10):
# Resize image to 10x10
image = cv2.resize(image, (dim, dim))
# Convert color space to LAB format and extract L channel
L, A, B = cv2.split(cv2.cvtColor(image, cv2.COLOR_BGR2LAB))
# Normalize L channel by dividing all pixel values with maximum pixel value
L = L/np.max(L)
# Return True if mean is greater than thresh else False
return np.mean(L)
def changeBrightness(image): if (isbright(image) < 0.5):
gammaImg = gammaCorrection(image, 1)
elif (isbright(image) > 0.85):
gammaImg = gammaCorrection(image, 0.75)
gammaImg = image
return gammaImg
def improve_brightness(overall_images):
new_arr = []
for i in range(len(overall_images)):
col = dict[i]
new_array_1 = []
for j in range(len(overall_images[i])):
img = changeBrightness(overall_images[i][j])
#appending the image to a new array
return new_arr

Contrast Adjustments

  • Adjusts image contrast by its histogram.
  • To enhance contrast, spreads out intensity range of image.
  • This allows the image’s areas with lower contrast to gain a higher contrast.
# import the neccessasry library
from skimage.exposure import is_low_contrast
def histogram_equalization(img_in):
# segregate color streams
b,g,r = cv2.split(img_in)
h_b, bin_b = np.histogram(b.flatten(), 256, [0, 256])
h_g, bin_g = np.histogram(g.flatten(), 256, [0, 256])
h_r, bin_r = np.histogram(r.flatten(), 256, [0, 256])
# calculate cdf
cdf_b = np.cumsum(h_b)
cdf_g = np.cumsum(h_g)
cdf_r = np.cumsum(h_r)

# mask all pixels with value=0 and replace it with mean of the pixel values
cdf_m_b =,0)
cdf_m_b = (cdf_m_b - cdf_m_b.min())*255/(cdf_m_b.max()-cdf_m_b.min())
cdf_final_b =,0).astype('uint8')

cdf_m_g =,0)
cdf_m_g = (cdf_m_g - cdf_m_g.min())*255/(cdf_m_g.max()-cdf_m_g.min())
cdf_final_g =,0).astype('uint8')
cdf_m_r =,0)
cdf_m_r = (cdf_m_r - cdf_m_r.min())*255/(cdf_m_r.max()-cdf_m_r.min())
cdf_final_r =,0).astype('uint8')

# merge the images in the three channels
img_b = cdf_final_b[b]
img_g = cdf_final_g[g]
img_r = cdf_final_r[r]

img_out = cv2.merge((img_b, img_g, img_r))
# validation
equ_b = cv2.equalizeHist(b)
equ_g = cv2.equalizeHist(g)
equ_r = cv2.equalizeHist(r)
equ = cv2.merge((equ_b, equ_g, equ_r))

return img_out

def improve_contrast(overall_images):
new_arr = []
for i in range(len(overall_images)):
col = dict[i]
new_array_1 = []
for j in range(len(overall_images[i])):

img = overall_images[i][j]
if(is_low_contrast(img, fraction_threshold=0.05, lower_percentile=1, upper_percentile=99, method='linear')):
img = histogram_equalization(img)
#appending the image to a new array
return new_arr
newarray = improve_contrast(newarray)

Sharpening Images

  • Edge detector used to compute the second derivatives of an image.
  • This determines if a change in adjacent pixel values is from an edge or continuous progression. Laplacian filter kernels usually contain negative values in a cross pattern, centered within the array. The corners are either zero or positive values. The center value can be either negative or positive.
def sharpenimage(image):
laplacian_var = cv2.Laplacian(image, cv2.CV_64F).var()
if laplacian_var < 100:
kernel = np.array([[0, -1, 0],
[-1, 5,-1],
[0, -1, 0]])
sharpened_img = cv2.filter2D(src=image, ddepth=-1, kernel=kernel)
sharpened_img = image
return sharpened_img
def improve_sharpening(overall_images):
new_arr = []
for i in range(len(overall_images)):
col = dict[i]
new_array_1 = []
for j in range(len(overall_images[i])):
img = sharpenimage(overall_images[i][j])
#appending the image to a new array
return new_arr

Body Segmentation

  • Media Pipe Segmentation function is used to blur the background of the image
  • The mask has the same width and height as the input image, and contains values in [0.0, 1.0] where 1.0 and 0.0 indicate “human” and “background” pixel respectively.

Pose Landmarks

  • Media pipe blaze pose is used to extract 3D coordinates of 33 joints from the image
  • x and y: Landmark coordinates normalized to [0.0, 1.0] by the image width and height respectively.
  • z: Represents the landmark depth with the depth at the midpoint of hips being the origin, and the smaller the value the closer the landmark is to the camera.

Angle Computation

Key angles at ( knee , elbow , shoulder , ankle ) are calculated from the points extracted and labelled with respective Asana name Angle at a joint is given by:

angle = degrees(atan2(y3 — y2, x3 — x2) — math.atan2(y1 — y2, x1 — x2))

def calculateAngle(landmark1, landmark2, landmark3): 
x1, y1, _ = landmark1
x2, y2, _ = landmark2
x3, y3, _ = landmark3

angle = math.degrees(math.atan2(y3 - y2, x3 - x2) - math.atan2(y1 - y2, x1 - x2))

# Check if the angle is less than zero.
if angle< 0:

# Add 360 to the found angle.
angle += 360

return angle

ML model results:

Train and test machine learning algorithms (Random Forest, SVC, Decision Tree, KNN, Adaboost, RFC) using the dataframe (csv) generated to find which model best fits.




Love podcasts or audiobooks? Learn on the go with our new app.

Machine Learning — How it works

How did I start learning ML? — in Nepal

Kaggle Titanic Challenge: Create New Features Using Extracted Data

Fruit and Vegetable Detection and Feature Extraction using Instance Segmentation-Part 2

Logistic Regression and Machine Learning

Supervised Learning for Dummies!

Do You Want To Know How Perceptron Algorithm works Internally

Part 3: Model Testing — Image Classification with YonoHub & Tensorflow V2.0 Series

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store


More from Medium

How to enable TensorFlow to use GPU drivers on Windows

How to analyze tweets?

The many uses of a Vector Database

SGPN paper explained