mxnet.image.imread can not correctly read jpg with orientation

[ksellesk]

See also:
#13216

This is a bug must be fixed since many jpegs has orientation.

People believe that mxnet will do right thing but waste time for debugging other thing.

[mxnet-label-bot]

Hey, this is the MXNet Label Bot.
Thank you for submitting the issue! I will try and suggest some labels so that the appropriate MXNet community members can help resolve it.
Here are my recommended labels: Bug, Feature

[ksellesk]

Such as this pic

[vdantu]

@mxnet-label-bot add [question, bug]

[anirudhacharya]

@ksellesk this is the doc for opencv's imread - https://docs.opencv.org/3.0-beta/modules/imgcodecs/doc/reading_and_writing_images.html?highlight=imread#cv2.imread

can you please provide a code example or usage of reading jpg with orientation. The method signatures of mxnet's imread and opencv's imread seem pretty similar.

[ksellesk]

see this one

https://docs.opencv.org/3.4.6/d4/da8/group__imgcodecs.html

https://github.com/apache/incubator-mxnet/blob/master/src/io/image_io.cc#L159

why IGNORE ORIENTATION?

[anirudhacharya]

b = mx.img.imread("img.jpg", flag=cv2.IMREAD_GRAYSCALE)

in the above code snippet except for cv2.IMREAD_GRAYSCALE and cv2.IMREAD_COLOR all the flags mentioned below errors out in mxnet while they are supported in opencv.

cv2.IMREAD_UNCHANGED
cv2.IMREAD_GRAYSCALE
cv2.IMREAD_COLOR
cv2.IMREAD_ANYDEPTH
cv2.IMREAD_ANYCOLOR
cv2.IMREAD_LOAD_GDAL
cv2.IMREAD_REDUCED_GRAYSCALE_2
cv2.IMREAD_REDUCED_COLOR_2
cv2.IMREAD_REDUCED_GRAYSCALE_4
cv2.IMREAD_REDUCED_COLOR_4
cv2.IMREAD_REDUCED_GRAYSCALE_8
cv2.IMREAD_REDUCED_COLOR_8
cv2.IMREAD_IGNORE_ORIENTATION 

[ksellesk]

img = cv2.imread("test.jpg")
img.shape
(4032, 3024, 3)
img = mx.image.imread("test.jpg")
img.shape
(3024, 4032, 3)

you can test this pic:
https://user-images.githubusercontent.com/9996886/57754105-47df8300-7720-11e9-94f0-7d9f1ec8766d.JPG

[anirudhacharya]

here is a weird discrepancy I am facing when using opencv with python and with c++ -

cv2.imread("img.jpg", flags=cv2.IMREAD_REDUCED_GRAYSCALE_2)

returns an image of shape (112, 112)
but when I use imread in c++ with

cv::imdecode(buf, cv::IMREAD_REDUCED_GRAYSCALE_2)

returns a matrix of shape (224,224,1).

The original image shape is (224,224,3)

[VinceJnz]

I'm having the same problem
mx.image.imread crashes when reading portrate oriented images.
I am using mxnet 1.7.0

Has there been any fix for this?