search

xtudbxk / DSRG-tensorflow

a tensorflow version for DSRG (Weakly-Supervised Semantic Segmentation Network with Deep Seeded Region Growing)
75 stars 20 forks source link

Question #1

Closed outsidercsy closed 6 years ago

outsidercsy commented 6 years ago

How could I get localization_cues-cal.pickle and init.model？ @xtudbxk

xtudbxk commented 6 years ago

Sorry for the late reply. You can refer to official website for DSRG (https://github.com/speedinghzl/DSRG), and download the localization_cues-cal.pickle and the init.model from the "Training the DSRG model" chapter.

outsidercsy commented 6 years ago

Thanks.But I find that the init.model is a caffe model file.When I directly use it in your tensorflow version code, an error turns out that the init.model can not be recognized as a tensorflow model file. Is there anything I should do with the init.model before using it?

xtudbxk commented 6 years ago

Yes, it is. You should use the convert.py to convert it from caffe model to npy file. And I just update the readme.md to provide the link for the converted model which you can download it directly. And for more details about the convert.py, you can refer to the project in https://github.com/xtudbxk/convert_caffemodel_to_npy .

outsidercsy commented 6 years ago

Get it. Thanks.

kevinlee9 commented 5 years ago

Hi, @xtudbxk . I find layers in the init caffe model(https://github.com/speedinghzl/DSRG) is different from the one provided in your repo(https://github.com/xtudbxk/DSRG-tensorflow).

There are some extra layers in vgg16_deeplab_aspp.npy, like fc6_2, fc7_2, fc8-SEC2, fc6_3, fc7_3, fc8-SEC3, fc6_4, fc7_4, fc8-SEC4, but how to get parameters of these layers? As pretrained vgg only has fc6, fc7 and fc8

xtudbxk commented 5 years ago
  1. It seems the network in origin DSRG project also use the deeplab-aspp as the baseline network. So I think the two project has the same layers.
  2. All my pretrained model is from the Deeplab project, I just convert all of it into npy file format. You can find the convert code from this project.
  3. And what's more, during my training, I find that even without aspp strcture, the best result of dsrg also reaches 57.0% miou. So you also can remove the aspp structure, and just use the pretrained model of deeplab-largefov.
kevinlee9 commented 5 years ago

@xtudbxk thanks, so vgg16_deeplab_aspp.npy is converted using this link http://liangchiehchen.com/projects/DeepLabv2_vgg.html?

xtudbxk commented 5 years ago

Yes, it is.