wangleihitcs/deeplab-v1-pytorch

Code for ICLR 2015 deeplab-v1 paper "Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs"

deeplab-v1-pytorch's Issues

学长好

我想问一下 py3.7的环境怎么装pydensecrf呢？
py3.5环境能装densecrf，但是torch和opencv又会有问题。
py3.7环境下对应的torch和opencv版本合适，但是又会出现No module named 'pydensecrf.densecrf'的情况

Question about "deeplab_largeFOV.pth"

I want to know the source of "deeplab_largeFOV.pth" in data directory, can you explain detail about is??

Is it a pre-trained weight on the ImageNet dataset?
or Is it a pre-trained weight on DeepLab-VGG16-LargeFOV?

I hope your reply. Thanks.

How to calculate the parameters of Pairwise potentials

Thank you for sharing your code! Your work is really brilliant!

But here I'm puzzled about how to set the parameters of Pairwise potentials with other datasets. I noticed that your parameters were set with im.shape == (513,513,3) :

post_processor = crf.DenseCRF( iter_max=10, # 10 pos_xy_std=3, # 3 pos_w=3, # 3 bi_xy_std=140, # 121, 140 bi_rgb_std=5, # 5, 5 bi_w=5, # 4, 5 )

I learned that the default parameters are set as below when im.shape == (640,480,3):

d.addPairwiseGaussian(sxy=3, compat=3) d.addPairwiseBilateral(sxy=80, srgb=13, rgbim=im, compat=10)

Generally, do I need to do mathematical derivation or I can just set the parameters according to some simple rules. Waiting for your opinion if you are free, thank you~

VOCdevkit

Why the folder of VOCdevkit have only two folder,SegmentationClassAug and __MACOSX

为什么nn.MaxPool2d后面接nn.AvgPool2d

如题。
这样有什么好处，为啥用两个，一个MaxPool2d不行吗

miou for 1464images