A quick python code to detect color of an item from ecommerce website using just kmeans without OpenCV. Nowhere close to accurate but good for a quick hack. Can be improved with Sobel edge detection and skin detection.

swatching.py

import numpy as np
from PIL import Image, ImageDraw
import colorsys, random
from pylab import plot,show
from numpy import vstack,array
from numpy.random import rand
from scipy.cluster.vq import kmeans,vq
from mpl_toolkits.mplot3d import Axes3D
import matplotlib.pyplot as plt
import matplotlib.image as mpimg


res = 2 # Resolution
clusters = 4

import glob, os
os.chdir("./images")
for file in glob.glob("*.jpg"):
    print(file)
files = glob.glob("*.jpg")
files.append(glob.glob("*.jpeg"))
filename = files[random.randint(0,len(files))]

im = Image.open("./" + filename)

im = im.crop((int(im.width*0.13), 0, int(im.width*0.87), im.height)) # crop the left/right bands that appear in some images to make bg removal easy
orig = im

##### Background removal code ########
border_clusters = 2
border_pixels = []
for i in  range(0, im.width):
	px = im.getpixel((i, 0))
	border_pixels.append([px[0],px[1],px[2]])
	px = im.getpixel((i, im.height-1))
	border_pixels.append([px[0],px[1],px[2]])

for i in range(0, im.height):
	px = im.getpixel((0, i))
	border_pixels.append([px[0],px[1],px[2]])
	px = im.getpixel((im.width-1, i))
	border_pixels.append([px[0],px[1],px[2]])


border_data = array(border_pixels)
# computing K-Means with K = 2 (2 clusters)
border_centroids,_ = kmeans(border_data, border_clusters)
# assign each sample to a cluster
border_idx,_ = vq(border_data, border_centroids)
border_counts = []
for i in range(0, border_clusters):
	border_counts.append(border_data[border_idx==i,1].size)

bg = border_counts.index(max(border_counts))

print border_centroids

bgR = border_data[border_idx==bg,0]
bgG = border_data[border_idx==bg,1]
bgB = border_data[border_idx==bg,2]


bgHSV = []

for i in range(0,bgR.size):
	bgHSV.append(colorsys.rgb_to_hsv(float(bgR[i])/255, float(bgG[i])/255, float(bgB[i])/255))
	print "hsv", bgHSV[i]


hRange = []
sRange = []
vRange = []

for i in bgHSV:
	hRange.append(i[0])
	sRange.append(i[1])
	vRange.append(i[2])

hMax = max(hRange)
hMin = min(hRange)
sMax = max(sRange)
sMin = min(sRange)
vMax = max(vRange)
vMin = min(vRange)



for i in range(0,im.width):
	for j in range(0, im.height):
		px = im.getpixel((i,j))
		pxhsv = colorsys.rgb_to_hsv(float(px[0])/255, float(px[1])/255, float(px[2])/255)
		if (hMin <= pxhsv[0] <= hMax) and (sMin <=pxhsv[1] <= sMax) and (vMin <=pxhsv[2] <= vMax):
			im.putpixel((i,j),(0, 255, 0, 255))


##### background removal code  END########

x = []
y = []
xy = []
xyz = []

# im = im.crop((int(im.width*0.333), int(im.height*0.333), int(im.width*0.6666), int(im.height*0.6666)))
im = im.crop((int(im.width*0.2), 0, int(im.width*0.8), im.height))
# im.show()


for i in range(0,im.width/res):
	for j in range(0, im.height/res):
		px = im.getpixel((i*res,j*res))
		if px != (0, 255, 0): #if the pixel is of key color don't insert it in data
			xyz.append([px[0],px[1],px[2]])

data = array(xyz)

# computing K-Means with K = 2 (2 clusters)
centroids,_ = kmeans(data,clusters)
# assign each sample to a cluster
idx,_ = vq(data,centroids)

for centroid in centroids:
	print centroid

counts = [] # count number of pixel per cluster
for i in range(0,clusters):
	counts.append(data[idx==i,1].size)


swatches =[]
for i in range(0, clusters):
	cen = centroids[i]
	swatches.append({'count': counts[i], 'color': (cen[0], cen[1], cen[2])})

# sorting the swatch according to the counts, the last item in the list is the color
sorted_swatches = sorted(swatches, key=lambda k: k['count']) 	
	
print swatches
print sorted_swatches

sw = Image.new('RGB', (50*len(swatches),50), (191,191,191))
dr = ImageDraw.Draw(sw)

itr = 0
colors = []
for centroid in centroids:
	colors.append([centroid[0], centroid[1], centroid[2]])
	itr+=1


for i in range(len(swatches)):
	f = sorted_swatches[i]['color']
	dr.rectangle(((i*50,0),(i*50+50,50)), fill=(f))

print 'len swatch', len(swatches)
dr.rectangle((((len(swatches)-1)*50,0),((len(swatches)-1)*50+10,10)), fill='Green', outline='white')
dr.rectangle((((len(swatches)-2)*50,0),((len(swatches)-2)*50+10,10)), fill='Red', outline='white')


##### Plotting the image using matplotlib #######
fig = plt.figure()
ax = fig.add_subplot(121, projection='3d')

for i in range(0,clusters):
	r = float(colors[i][0])/255
	g = float(colors[i][1])/255
	b = float(colors[i][2])/255
	ax.scatter(data[idx==i,0],data[idx==i,1], data[idx==i,2], c=array([[ r],[ g],[ b]]), marker='o')

for i in range(0, clusters):
	r = float(colors[i][0])/255
	g = float(colors[i][1])/255
	b = float(colors[i][2])/255
	ax.scatter(colors[i][0],colors[i][1],colors[i][2], c=array([[ r],[ g],[ b]]), marker='*', s = 150, linewidths=2)


ax.set_xlabel('Red')
ax.set_ylabel('Blue')
ax.set_zlabel('Green')

dx = fig.add_subplot(322)
dx.axis('off')
dx.imshow(array(orig))

cx = fig.add_subplot(324)
cx.axis('off')
cx.imshow(array(im))

bx = fig.add_subplot(326)
bx.axis('off')
bx.imshow(array(sw))

plt.show()