Do you have any details on how to train a Classification model for person attributes. I want to seek multiple attributes of a particular person eg wearing hat, glasses etc. I have prepared a dataset of 1000 person images with 4 binary attributes.
Firstly you need to prepare the images for each class.
Then use TLT classification network to train.
For reference, you can try to run classification notebook inside TLT 2.0_dp docker.
Images have already been prepared. I need help on how to arrange them.
Eg photo 1 has male, wearing hat, short hair, wearing glasses (all values set as 1)
Photo2 has male, wearing hat, short hair, wearing glasses (1,0,0,0)
How do I arrange this to trigger training.
What will be the inference output? Will it be top-k classes?
How many classes do you want to train? You need to crop and split your images to corresponding class.
Currently, tlt classification just shows top-1 class. But you can see the all prediction value in the log.
Around 4 classes. That would mean that the same image would be present in multiple Attribute folders depending on the applicability of the attribute.
Can you refer to VOC dataset(also mentioned in jupyter notebook of the docker) ? You can take a look at the public VOC dataset and inspect how to split the image.