Image Pre-processing about alignedreid-re-production-pytorch

[huanghoujing]

The processing steps I used:

• resizing to w x h = 128 x 256
• subtracting image mean
• dividing by image std
• horizontally flipping with probability 0.5

No cropping is used. Though the code provides the option to use cropping, I found it did not benefit performance in my experiments.

[Phoebe-star]

do you have use resnet or googlenet or VGG .... with tensorflow and it is trainable for imagenet

[huanghoujing]

I only use ResNet-50 here. I once tried Inception-V3 for identification loss and it's worse than ResNet-50.

[Phoebe-star]

I find use the trainable resnet 50 ,if I set the batch size 128 , it express 128x3 images, because triplet.
the network will out of memory.

[huanghoujing]

128 triplets are formed among this batch of 128 images, so only 128 images are needed and the memory is enough.