a question about the implementation of OCRNet in MMSegmentation

[finestprogrammer]

Hi contributors, thank you for your awesome project. I have a question about the implementation of OCRNet. As I read in the original paper, the OCRNet pipeline contains a module called Soft Object Regions and a module called Pixel Representations. Also, the OCRNet in MMSegmentation is implemented as a cascaded encoder-decoder with a FPN head and a OCR head followed by OCRNet. Are Soft Object Regions and Pixel Representations produced by applying FPN on outputs of HRNet? Please point out if I misunderstood anything?

[xiexinch]

Hi @createhappily,
The pixel representation is the output of the backbone and the soft object regions is produced by the FCN head.

[finestprogrammer]

Thank you.