I’m training with a custom dataset on a DetectnetV2 architecture with ResNet-34 backbone. My dataset contains about 15K RGB images along with their respective labels in KITTI format. I have calculated per-channel mean values and the standard deviation for my whole dataset and want to normalize each image as a pre-processing step before training pass.
I see that FasterRCNN has
network_config to support normalization. I couldn’t find similar configuration for DetectNetV2.
Also, I see that DetectNetV2 only accepts images of the format ‘jpg’, ‘png’, and ‘jpeg’. This eliminates the possibility of generating a dataset with images having floating point pixel values as all the three aforementioned formats hold uint pixel values.
Is there a way I could utilize per-channel mean values and standard-deviation of my dataset for training DetectnetV2 model by adjusting something from experimentation specification file for training? If not, is there a way to save normalized images with floating point pixel values between 0 and 1 in jpg/png/jpeg formats?