This question is strongly connected to: http://answers.opencv.org/question/178680/how-to-use-meanbinaryproto-with-blobfromimages/
I'm working with OpenCV 3.4.1 and caffe 1.0.0
I have different results for classification in caffe and OpenCV only when I'm using mean subtraction, without it everything works fine. I have the same observation for some caffenet and squeezenet_v1.1 models. My code:
import cv2
import caffe
import numpy as np
# function comparing results of image classification in opencv and caffe:
# with mean substraction results don't match
def caffe_vs_ocv_with_mean_subs(caffemodel, deploy_prototxt):
# read mean
blob = caffe.proto.caffe_pb2.BlobProto()
with open("mean.binaryproto", 'rb') as f:
blob.ParseFromString(f.read())
data = np.array(blob.data).reshape([blob.channels, blob.height, blob.width])
cv_mean = [np.mean(data[0]), np.mean(data[1]), np.mean(data[2])]
# define caffe net
net = caffe.Net(deploy_prototxt, caffemodel, caffe.TEST)
transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
transformer.set_mean('data', data)
transformer.set_transpose('data', (2, 0, 1))
# get caffe prediction
img = cv2.imread("example.jpg")
img = cv2.resize(img, (227, 227), cv2.INTER_LINEAR)
net.blobs['data'].data[...] = transformer.preprocess('data', img)
out = net.forward()
print "mean: " + str(cv_mean)
print "caffe probs: " + str(out['prob'][0])
# define openCv net
cv_net = cv2.dnn.readNetFromCaffe(deploy_prototxt, caffemodel)
# get opencv predictions:
img_for_cv = cv2.imread("example.jpg")
# swapRB=False, just like in http://answers.opencv.org/question/178680/how-to-use-meanbinaryproto-with-blobfromimages/,
# but I tried any combinations of swapRB, crop flags
cv_blob = cv2.dnn.blobFromImage(img_for_cv, 1, (227,227), cv_mean, swapRB=False, crop=False)
cv_net.setInput(cv_blob)
prob = cv_net.forward()
print "cv probs" + str(prob)
# no mean substrucation - this works fine
def caffe_vs_ocv_simple(caffemodel, deploy_prototxt):
# no mean substraction
# define caffe net
net = caffe.Net(deploy_prototxt, caffemodel, caffe.TEST)
transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
transformer.set_transpose('data', (2, 0, 1))
# get caffe prediction
img = cv2.imread("example.jpg")
img = cv2.resize(img, (227, 227), cv2.INTER_LINEAR) # the same way as resizing cv2.dnn.blobFromImage when crop=False
net.blobs['data'].data[...] = transformer.preprocess('data', img)
out = net.forward()
print "caffe probs: " + str(out['prob'][0])
# define openCv net
cv_net = cv2.dnn.readNetFromCaffe(deploy_prototxt, caffemodel)
# get opencv predictions:
img_for_cv = cv2.imread("example.jpg")
cv_blob = cv2.dnn.blobFromImage(img_for_cv, 1, (227,227), (), swapRB=False, crop=False)
cv_net.setInput(cv_blob)
prob = cv_net.forward()
print "cv probs" + str(prob)
# tested on different models:
# normal reference caffemodel (but with some other mean):
caffe_vs_ocv_with_mean_subs("caffenet_1000.caffemodel", "caffenet_deploy_1000.prototxt")
# my custom versions of caffenet and squeezenet with different number of output classes
caffe_vs_ocv_with_mean_subs("caffenet_5.caffemodel", "caffenet_deploy_5.prototxt")
caffe_vs_ocv_with_mean_subs("squeezenet_1.1.caffemodel", "squeezenet_deploy.prototxt")
# all tests with mean substraction failed, giving completely wrong results
# examples without mean substraction:
# here differences between probabilities are tolerable, it works fine:
caffe_vs_ocv_simple("caffenet_model.caffemodel", "caffenet_deploy.prototxt")
caffe_vs_ocv_simple("caffenet_5.caffemodel", "caffenet_deploy_5.prototxt")
