Closed youkaichao closed 6 years ago
They are mean values of R/B/G relatively. They are from the mean values of ImageNet dataset. Since the pretrained model is converted from Caffe with caffe-tensorflow, if you are feeding a natural image to the pretrained model, you need to normalize it (i.e. subtract the mean values) to get desired output.
However, because we are feeding "white rendered images over black background (0's) ", in my experiments, there is no need to subtract the values.
I'm closing this. Feel free to open this again if you find any further issue.
Where does the magic numbers of 104., 116., 122. in input.py/Shape/subtract_mean come from ? Does these numbers have special meaning?