search

juntang-zhuang / ShelfNet

implementation for paper "ShelfNet for fast semantic segmentation"
MIT License
252 stars 41 forks source link

The inference speed on Cityscapes #2

Closed githubfragments closed 5 years ago

githubfragments commented 5 years ago

Nice work! I wonder how much time you speed to process the images of Cityscapes. Thus we can compare it with ICNet.

juntang-zhuang commented 5 years ago

Thanks for your interest. The code to test running speed is in experiments/segmentation/test_speed.py.

I test speed with a single random tensor of size (1,3,512,512), repeat and average the running time. In this way, the running speed is related to the size of the image, but irrelevant to which dataset.

I did not re-implement ICNet in PyTorch and did not compare with it. But it can be tested with test_speed.py if the model is well defined in PyTorch.