The shape of tensor obtained by nvidia.dali reads the image is rotated

[miaomiaoma0703]

The shape of tensor obtained by nvidia.dali reads the image is rotated, because the rotated jpeg file has orientation info, dali considers the rotation head information of the jpeg file, but opencv does not consider it。 code: dali：

class ImageDecoderPipeline(Pipeline):
    def __init__(self, image_dir, batch_size, num_threads, device_id):
        super(ImageDecoderPipeline, self).__init__(batch_size, num_threads, device_id, seed = seed)
        self.input = ops.FileReader(file_root = image_dir)
        if torch.cuda.is_available():
            self.decode = ops.ImageDecoder(device = 'mixed', output_type = types.RGB)  #mixed
            # self.rotate = ops.Rotate(device="gpu", angle = 0.0)
        else:
            self.decode = ops.ImageDecoder(device='cpu', output_type=types.RGB)  # mixed

    def define_graph(self):
        jpegs, labels = self.input()
        images = self.decode(jpegs)
        return (images, labels)

opencv： input = cv2.imread(file_path)

the jpeg file:

the head info: {"size":2534018,"format":"jpeg","width":3000,"height":4000,"colorModel":"ycbcr","orientation":"Right-top"}

obtain by dali and opencv： dali: opencv:

[JanuszL]

Hi, Orientation is just metadata that is used by viewers to rotate the image before the displaying, but the image is encoded as it is. DALI doesn't use metadata to adjust the image after the decoding. If you want to apply the Exif orientation to the actual image I would recommend using imagemagic with -auto-orient option to adjust your image first and then pass it to DALI.

[miaomiaoma0703]

Hi, The fact is that the part of the image decoded with the dali will be rotated (not the same as the one opened with the viewing software), but opencv will not cause the rotation. I suspect it is a bug in the dali library. The dali decoding code as above.

[JanuszL]

Both viewers and OpenCV apply optional rotation operation based on the Exif metadata - https://github.com/opencv/opencv/blob/master/modules/imgcodecs/src/loadsave.cpp#L912. In the case of DALI, we are not doing such an operation. We will add it to our ToDo list, and we will be more than happy to review and accept any PR introducing such change from the community.

[mohanksriram]

Hi @JanuszL, As you suggested, I will be starting work on this issue. OpenCV has its own parser file here: https://github.com/opencv/opencv/blob/master/modules/imgcodecs/src/exif.hpp I have two questions:

• What points should I keep in mind when adapting exif parser for DALI?
• Where should the parser file be added? (In the dali/image module?)

[JanuszL]

Hi @mohankumarSriram:

• I would make sure that there is an additional option in the image.decoder that allows the user to decide if apply EXIF metadata to the images
• also please check what images may have EXIF meta attached - I think TIFF also supports it
• add a dedicated code that would flip the image on demand. DALI has flip kernels/operators you can reuse and just call if needed
• dali/image sounds good