Train data for Semantic segmentation using existing Nets (e,g.Segnet) for different classes

There is an example in Matlab "Semantic Segmentation Using Deep Learning" which is used to train objects in classes which are already existing in Original segnet. How can I use SegNet for indoor scene understanding which may have completely different classes.

[ahmed anis]

i train data using the code in examples but after training when i change the do training to false it shows me an error..

[ahmed anis]

it says data = load(pretrainedSegNet); is not a defined variable

Answer 1

[Anh Tran]

The example starts with training using VGG-16 architectures and weights, rather than starting from scratch. VGG-16 is used for a wide range of classification (1000 classes) and it does not limit to outdoor. I understand it as SegNet does segmentation by integrating VGG-16 features extraction and classification abilities. Therefore, you may want to provide indoor training images and follow the example to create and train the network. You can also explore different pretrained network and find the right one for your application. Hope this helps.