Segmentation accuracy is better during training than testing for validation (val) images

[briangow]

I observe with my image set (1 class + background) that I achieve much better segmentation accuracy for the validation (val) images during training (pixelAcc = 97%) then during testing (pixelAcc=89%). By default fcnTrain.m and fcnTest.m are setup to use the same val images. My understanding is that testing should actually be done on a fully independent (3rd) set of images but when running the default setup I would expect the same accuracy from fcnTrain and fcnTest for a given epoch. I confirmed by outputting images from predictions in SegmentationAccuracy.m that the segmentations for the val images look better during fcnTrain then fcnTest.

[ibcny]

I guess you should set the random seed as stated in fcnTrain.m based on epoch and the random seed used in training to successfully regenerate the results previously produced like:

rng(epoch+opts.randomseed);