Aws-cli: Pass image bytes to Rekognition without S3 not working

According to the documentation, one can pass image bytes to a Rekognition API operation by using the Bytes property in the --image option:

aws rekognition detect-labels --image "Bytes=blob"

I tried countless combinations, but to no avail. I also understand, that the image should be base64 encoded. However, whatever I try, I end up with the following error message.

An error occurred (InvalidImageFormatException) when calling the DetectLabels
operation: Invalid image encoding

It seems that the aws-cli takes the input "blob" you give it, and base64-encodes it again. The problem is, it doesn't in fact treat the input as a blob. It treats it as a string and tries to interpret it as utf8, so you can't actually pass a blob unless by coincidence you happen to be in possession of a blob that just happens to be comprised of entirely valid utf8 characters. Even if you pass JSON, it still encodes the Bytes value to base64.

As a consequence, the only option is to upload the image to S3 first, and then refer to it by using the S3Object property. However this is not always desired and the lack of "Bytes=blob" support seems to be in contrast to the documentation and build-in help texts.

System/software used:
aws-cli/1.11.175 Python/2.7.13 Linux/4.8.0-2-amd64 botocore/1.7.33