Video to Image Regression

Hello!

I have 32x32x256 (HeightxWidthXFrames) greyscale video data that I need to regress to a 32x32 image.

1. What is the ideal format for me to save my input data so it can be read into a NN? Is there an appropriate image format? (I have not been succesfull using .mat files in an image datastore)
2. Should I use a 2d or 3d ImageInputLayer? I intend to use a Unet architecture.

Thank you!

Answer 1

[Shashank Gupta]

Hi Michael,

Since your input to the model is a video data, it is appropriate to use 3D image datastore. Also Unet archtecture you intent to design will be a 3D architecture and in that case going for 3d imageInputLayer is prompt.

When we deal with high dimension data, It is always good choice to go with ".mat" data storage. In particularly your case, you can write a custom function in @ReadFcn property of datastore to read the ".mat" file.

I hope this helps you,

[Michael Keeling]

Thank you!