For Caffe framework, its pre-processing for model is fixed. Vitis AI offers several
pre-optimized routines like dpuSetInputImage()
and
dpuSetInputImageWithScale()
to perform image pre-processing on CPU
side, such as image scaling, normalization and quantization, and then data is fed into
DPU for further processing. These routines exist within the package of Vitis AI samples.
Refer to the source code of DNNDK sample ResNet-50 for more details about them.