Why is the downsize 16?

[daimaohui]

Can we set the downsize to 32?

What's the difference between 32 32 and 16 16？

run_demo.py superpixel grid cell, must be same as training setting why?

Can we use 16 16 training 32 32 to predict?

[fuy34]

Hi, Thanks for your interest in our work. You can do 32 for sure. We use 16 just for a better trade-off between speed and performance.

if you wish to use 3232, you probably need to retrain the network, as the provided checkpoint are trained with 1616.

[daimaohui]

Can we use 16 16 training，32 32 to predict?

[daimaohui]

In my own dataset, I use 16 16 for training and 16 16 for testing, and the final BP is 0.75. But I use 16 16 for training and 32 32 for testing, and the final BP is 0.88. However, I use 32 32 for training and 32 32 for testing, and the final BP is 0.44. All parameters of the three experiments are the same. So is it correct to use 16 16 for training and 32 32 for testing?

[fuy34]

This is interesting. We never tried to test with a different size than the training setting. I, personally, believe it is weird. Could you share some visualization? Also, how are the BR and other metrics? Looking into BP alone may cannot provide sufficient signals to study what is going on.