Hi, I am trying to understand your code. You code is very clean and well arranged

Hi, if you check our paper, we use different hyperpixel_ids fo

some questions. about cost-aggregation-transformers HOT 2 CLOSED

5100117 commented on June 18, 2024

some questions.

from cost-aggregation-transformers.

SunghwanHong commented on June 18, 2024 2

Hi,

if you check our paper, we use different hyperpixel_ids for different datasets. Specifically, hyperpixel_ids=[0,8,20,21,26,28,29,30] is for SPair-71k found by HPF (https://github.com/juhongm999/hpf)
Yes, we tried using other backbones (check out our ablation study on feature backbone), but we have not tried resnet152. Have you tried training from scratch when you replaced to resnet152? if you simply replaced without training, it's highly likely that it may perform worse. If it's not the case, there may be several reasons that could result worse performance: (1) as you mentioned, un-optimal hyperpixel_ids, (2) insufficient training epochs or could be anything else!

We are planning to improve both the paper and code in the near future, and as soon as they are ready, we will share with the community

from cost-aggregation-transformers.

5100117 commented on June 18, 2024

I get it. Thank you very much.
you are awesome!

from cost-aggregation-transformers.

Recommend Projects