search

microsoft / O-CNN

O-CNN: Octree-based Convolutional Neural Networks for 3D Shape Analysis
MIT License
729 stars 171 forks source link

TF Semantic Segmentation with UNet #103

Closed migari01 closed 4 years ago

migari01 commented 4 years ago

Hello,

I have been trying to run your code for semantic segmentation with UNet on the Bed and Bottle dataset you provided, and the network does not seem to converge. I have used the same configuration as HRNet config, but changed option model from "hrnet" to "unet". Did you use a different configuration for your experiments? If yes, could you please share it?

Also, is this network the equivalent of Caffe O-CNN version for segmentation? Are there any significant differences?

Thank you

jeichelbaum commented 4 years ago

If I understand both networks correctly, then one of the major differences is that the Tensorflow unet uses residual blocks and is potentially much deeper than the shallow Caffe segmentation network. The author will probably be able to give a more conclusive answer though.

wang-ps commented 4 years ago

Yes, the tensorflow-based UNet is deeper with resblocks and its performance is also better.
What do you mean by "does not seem to converge"? Can you run the experiments using HRNet with the instructions here?