I am trying to take binary masks say with two different people and everything else is black. Now I want to group each person into their own clusters using K-means so that I can eventually draw bounding boxes around them. Here is the code I have so far:
def kmeans(img):
k_values = range(1, 5)
pixels = np.float32(img.reshape(-1,1))
criteria = (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 10, 1.0)
flags = cv2.KMEANS_PP_CENTERS
min_ssd = 0
for k in k_values:
ssd,labels,centers = cv2.kmeans(pixels,k,None,criteria,10,flags)
if k == 1 or ssd < min_ssd: #looking for elbow in graph
min_ssd = ssd
min_labels = labels
min_centers = centers
return min_labels,min_centers
As it stands, this code is returning the centers as pixel values and not coordinates so I am losing the spatial information in the frame. I am not quite sure how to change this code in order to obtain the cluster center as a point in the image and be able to visualize the clusters.
Edit: I have changed the approach, first I start with OpenCV's simpleblobdetector() to detect the centroids of each blob in the image. From here I am initializing the centroids with those values. I plan to update the code when I finish making the changes
Final Code
def kmeans_helper(img, centroids):
data = np.nonzero(img)
data = tuple(zip(data[0],data[1]))
data = np.asarray(data).reshape(-1,2)
num_clusters = len(centroids)
kmeans = KMeans(n_clusters = num_clusters, init = centroids, n_init = 1)
kmeans.fit(data)
centroids_array = np.uint8(kmeans.cluster_centers_)
num_labels = np.unique(kmeans.labels_)
clusters = [data[kmeans.labels_ == label] for label in num_labels]
return clusters
I figured I would update my final code for anyone who might need this in the future.
(x,y)points should do the trick,kmeanswill return each cluster's centroid also as an(x,y)array, and that should give you the centroid of each blob. Side note: if you are looking just for each blob centroid, you could also carry out component analysis and calculate the bounding box of each blob, from there it should be fairly easy to compute the centroids. You can also use central moments to compute blob centroids.