awesome-semantic-segmentation-pytorch
awesome-semantic-segmentation-pytorch copied to clipboard
Tramac
About DFAnet
Did u test the inference speed of DFAnet? The result of my test by mxnet is over 40ms for one image(size:1024*1024), which is far away from the speed reported in the paper. And it's really hard to train DFAnet. The best mIOU I got is close to 50% on Cityscape....
I guess it is because that the paper result is pre-trained on ImageNet?
The first question is why it's so fast. I built the model with 40ms inference time, with regard to 10ms which is mentioned in the original paper.
I tested it with 2 1080ti.The result is nearly 50ms per frame.I use cudnn and pytorch.It seems a big margin between our model and original paper's model.
the poor implement of depthwise conv in pytorch lead to the much slower inference speed. @yellowYuga @Aktcob . But I don't know why the result is still poor.
of depthwise conv in pytorch lead to the much slower inference
MXNET is the same, actually~~~
where can I get the pre-trained model?
Did u test the inference speed of DFAnet? The result of my test by mxnet is over 40ms for one image(size:1024*1024), which is far away from the speed reported in the paper. And it's really hard to train DFAnet. The best mIOU I got is close to 50% on Cityscape....
where can I get the pre-trained model? please
Did u test the inference speed of DFAnet? The result of my test by mxnet is over 40ms for one image(size:1024*1024), which is far away from the speed reported in the paper. And it's really hard to train DFAnet. The best mIOU I got is close to 50% on Cityscape....
hi, @Aktcob, did you resize the cityscapes image into 10241024 or divide one 10242048 into two 1024*1024?
update; 53.8% mIoU....still far away...
Hi! May I ask what is the latest result of the miou of DFANet?
