search

vikolss / DACS

Code from the paper "DACS: Domain Adaptation via Cross-domain Mixed Sampling"
MIT License
123 stars 19 forks source link

Questions about the Crop Size #2

Closed qianyuzqy closed 4 years ago

qianyuzqy commented 4 years ago

Hi, thank you so much for your open-source code! It is well-organized and great!

I found that the crop size is (128, 512) in your config.json, while you mentioned that the images are randomly cropped to (512, 512) in your paper? I was wondering what is the final crop size you used to get the best performance of 52.14 in the final experiment?

Another question is that is it necessary to resize the cityscapes image from (h,w) to (h/2, h/w) and to resize the gtav image from (760,1280) to (300, 600) during the process of the random crop? It is not mentioned in your paper, will it affect the performance a lot?

Thank you very much!

All the best,

WilhelmT commented 4 years ago

Hi, thank you for your interest

You are right, the crop size used in the paper was indeed 512 and has been changed accordingly.

The Cityscapes images are resized to 512x1024 before cropping, and the gta images are resized to 760x1280 (same as synthia). This was consistent with existing work using DeepLabv2 for these benchmarks. We did not experiment with how different choices would affect performance.

qianyuzqy commented 4 years ago

Thanks for your reply! Best wishes to you!