Always returning similar softmax probability in deploy network.

[l...@inwatec.dk]

So I've made some changes to the CIFAR10 network in order to use my own dataset and to get a binary output (object present/object not present). In training it gets accuracies around 97% and that's confirmed in testing. However, when I try to use the cifar10_full.prototxt deploy network in my C++ code, regardless of whether the object is present or not I get an "object present" probability of between 0.071-0.075 (different for different images, but same result for each specific image when retested). If I change any of the layers in the network file the probabilities all go to 0.5.
What's happening here?

Thanks in advance