Why is the model deeplabv2?

[juajua56]

Can't we train with HRDA + ProCST model?

[shahaf1313]

Hey juajua :) We indeed haven't assembled to our architecture more accurate segmentation models, such as DAFormer and HRDA. It can potentially upgrade performance, but it also requires an enormous amount of resource usage such as RAM memory, and thus restricts the maximum output image size. Low resolution images are insufficient to achive improved performance on the UDA-SS task, therefore we haven't conducted those experiments. Best regards, Shahaf