search

junfu1115 / DANet

Dual Attention Network for Scene Segmentation (CVPR2019)
MIT License
2.41k stars 483 forks source link

evaluating speed is so slow #56

Open kaixinxiao opened 5 years ago

kaixinxiao commented 5 years ago

When I use test.py to evaluate model, the speed is so slow. Specifically, Costing 20 mins in single scale and 7 hours in multi-scale. Is it true?

yazhe2017 commented 5 years ago

Yes, my test process is also very slow. Hope author could explain about it.

junfu1115 commented 5 years ago

Due to evalute images with sliding the window, it would take a lot of time. Moreover, in multi-scale testing, we adopt several scales, and the largest scale is set to 2.2 times of original image, it might also take more time

kaixinxiao commented 5 years ago

@junfu1115 Thank you! But is it necessary to visualize the output using multi_scale? It costs dozens of hours but it has a better performance when I push it to cityscapes server. Do you apply multi_scale method about the performance of 81.5% on test set? Should I use the strategy with basesize=2048 and crop_size=1024?

xdsonglinliu commented 5 years ago

Hi, I have a question. Where can I see the sliding the window operation in the code when i test a large image(larger than the images used in training)? Anyone can help me? Thanks a lot! @kaixinxiao @yazhe2017

A-Kerim commented 2 years ago

Indeed it is very slow; any solutions?