Question about Segmentation tutorial

[orram]

Thanks for the project!
I noticed in your Tutorial: Class Activation Maps for Semantic Segmentation you mask the output of the network using the target class mask - this results in visualizing the contribution to the correct predictions by the model and not the prediction in general (in this case, why did the model decide these pixels are cars)

The mask can be important if querying about a particular pixel or group of pixels.

Thanks!

[caihuaiguang]

I think the using of different target classes is entirely dependent on the specific needs of the user. In general, using the model's prediction as the target class seems more natural in life. In my paper “CAMs as Shapley Value-based Explainers”, I chose to use the predicted label rather than the correct label to measure the performance of the explainability of different CAM methods.