search

Luffy03 / VoCo

[CVPR 2024] VoCo: A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis
Apache License 2.0
120 stars 9 forks source link

Details about this code #13

Open aifeixingdelv opened 3 months ago

aifeixingdelv commented 3 months ago

Hello，from the overall framework picture, I found the count of randomly crop is one. I want to know the method processes the random crops in a data picture in your model？ How to get more randomly crop from one picture.

Luffy03 commented 3 months ago

Hi sw_batch_size is used to define the number of random crops

aifeixingdelv commented 3 months ago

Hi sw_batch_size is used to define the number of random crops

Thank you for your rapid reply. I am curious about that if the part of the random crops can comprise the full picture?

Luffy03 commented 3 months ago

no it cannot

aifeixingdelv commented 3 months ago

Why? A little curious. I am not a researcher in medicine but in computer vision. I think your idea can been applied other dense prediction task if it can predict the full image.

Luffy03 commented 3 months ago

If it covers the whole picture It will share 100% area with all bases then we cannot generate diverse position labels

aifeixingdelv commented 3 months ago

If it covers the whole picture It will share 100% area with all bases then we cannot generate diverse position labels

So in term of medicine image segmentation, it doesn't need the full image prediction?

aifeixingdelv commented 3 months ago

If it covers the whole picture It will share 100% area with all bases then we cannot generate diverse position labels

I mean the whole crops not only one crop. If the whole crops can comprise the full image. Thank you for your reply.

Luffy03 commented 3 months ago

If it covers the whole picture It will share 100% area with all bases then we cannot generate diverse position labels

So in term of medicine image segmentation, it doesn't need the full image prediction?

not usual un training but it is required in inferrence

aifeixingdelv commented 3 months ago

If it covers the whole picture It will share 100% area with all bases then we cannot generate diverse position labels

So in term of medicine image segmentation, it doesn't need the full image prediction?

not usual un training but it is required in inferrence

Thank you. I read the paper carfully again and understand your idea!