Which model are you currently using? The only model that I've seen to work in grayscale is mnist.
Try this model and see if it works for you: https://github.com/BVLC/caffe/blob/master/examples/mnist/lenet.prototxt
You will need to change the input batch parameter from 61 to 1.
another question, does the mvnc sdk support a convolution kernel which is not square(such 3x1)? I've found that there is a convolution layer which include a kernel like this.
The Intel Neural Compute SDK should support non-squared. Could you share your model for me to test? You can post them here or send them over via private message, just click my user name and hit message.
Also, have you tried to use your model using the OpenVINO toolkit?
I'm have not tested mobilenet with a non-square input. Is there a particular reason why you are trying to use non-square inputs? I will need to do some test and get back to you.