Custom Dataset: Number of Classes

[bonlime]

Custom Dataset: Number of Classes

[bonlime]

It depends on your loss function and classes topology.
If classes are mutually exclusive you can add background class and use softmax at the top. During training at Imagenet authors of original paper used this way.
But you can use sigmoid as activation at the last layer without adding background and it also will work

[vodp]

As @bonlime said if you background contains other classes you may not want to assign it as a 'background' class, rather just ignore it as 'ignore_label'. Of course in detection the network segments every pixel so that you won't have background being classified.

If background is really irrelevant objects, then you may want to add it as an additional class.

In VOC pascal the Deeplabv3 used background class as it was defined in the dataset.
In city datasets, background class is absent because dense segmentation was designed for those datasets.

[vodp]

You need to write a custom loss function, addressing ignore_label and do not include the loss incurred by ignore_label into the total loss.

[ad12]

Thank you @bonlime and @vodp !

In the case that the pixels are mutually exclusive, would my prediction mask have 4 channels? Or would I make the mask non binary (i.e. each pixel has a value of 0-num_classes for the corresponding pixel value?

In other words, assuming my input image is of shape (288, 288, 1), then would my corresponding mask for 4 classes (not including background) have a shape of (288, 288, 1, 4) if I want the network to predict each mask separately?

[ad12]

@vodp I am planning on not adding a background class, but how do I ignore the pixel? Is there a way to specify it as ignore_label?

I am also not getting a binary mask after training. Are there any changes that I need to make for this?