search

martinkersner / train-DeepLab

Train DeepLab for Semantic Image Segmentation
MIT License
172 stars 76 forks source link

Doubt regarding deeplab for new dataset #10

Open srika91 opened 7 years ago

srika91 commented 7 years ago

If we want to use deeplab for new dataset that is substantially different from the trained datasets,how do we go about the training this network?

My understanding: deeplab=DCNN(trainable) + CRF(not trainable-but can cross validate parameters)

So we train deeplab for the new dataset,ie training DCNN and then cross-validate the CRF parameters.Is this correct or is it possible to train deeplab end-to-end ,meaning training CRF also?

hehebloom commented 7 years ago

Have you ever solved your problem? I try to train Deeplab on SUNRGBD dataset but it has 38 classes and looks quite different from VOC12, finally result I got looks terrible. Have I need to do something with the labels, or change some parameters in the code? If you have any hint it will be very pleasure.