search

daijifeng001 / MNC

Instance-aware Semantic Segmentation via Multi-task Network Cascades
Other
489 stars 182 forks source link

image size my own data #58

Open oneOfThePeople opened 7 years ago

oneOfThePeople commented 7 years ago

I know that the network resize shortest side to 600pix In my data set i get something like 5000x5000pix and i want to crop to fixed size There is advice for wich window size to choose? Something like 512x512? or 600x600 is better? Thank alot

hgaiser commented 7 years ago

There's a tradeoff that you need to make for the accuracy that you want and the computational resources that you want to spend on that. More pixels means more resources (memory-wise and time-wise), but you also get higher accuracy. In other words it is up to you ;)

oneOfThePeople commented 7 years ago

ok and there is way to prevent from scaling the image? thank you for the quick answer :)

hgaiser commented 7 years ago

Not through a parameter or something. You could set the parameter which scales to 600 at the moment to 512 or whatever your images already are. It will then scale to 512 but they already are 512x512 pixels so nothing will change. If you really want to disable scaling you'd have to modify the code a bit to do so.

oneOfThePeople commented 7 years ago

thank you.