Use scipy.misc.imread instead of PIL Image

[meetps]

WRT

https://github.com/warmspringwinds/tf-image-segmentation/blob/master/tf_image_segmentation/utils/tf_records.py#L37

I think you should reconsider using scipy.misc.imread in place of PIL Image, because for VOC datasets, the PIL loads a grayscale image for the segmentation annotation, while a 3 channel input is necessary to determine the segments.