I have 85 raw, unlabeled images for detecting my company’s logo using GluonCV. I plan on using AWS SageMaker to label and train an object detector, but I am not sure if/how to preprocess my images beforehand.
I have a training script that uses pre-trained weights from ssd_512_mobilenet1.0_custom and applies the SSDDefaultTrainTransform preprocessor.
# model model = gcv.model_zoo.get_model('ssd_512_mobilenet1.0_custom', classes=["logo"], pretrained_base=False, transfer='voc') # preprocessor train_dataset.transform(SSDDefaultTrainTransform(height = 512, width = 512))
My questions are as follows:
The SSDDefaultTrainTransform preprocessor “includes tons of image augmentations” according to the docs. I am not sure if this includes creating images to increase the size and diversity of the training data. If not, is there an API in GluonCV that can create additional images (either before or after labeling)?
For my application, colors are not important. I am only interested in detecting the shape of the logo. Should I apply a grayscale transformation to the images prior to SSDDefaultTrainTransform to simplify the images?