Seg-Fix and OCRnet question about openseg.pytorch HOT 4 CLOSED

We do not fully understand your first question as we apply the OCR block on the feature map output from the ResNet/stage-4 directly in the ocrnet.py file. We only apply bilinear-upsample on the final predicted logits map, which is the same as the previous DeepLab series and PSPNet series.

you apply the feature network after that use the OCR block after that you use the F.interpolat 2 times

As for the SegFix paper, we will update the news once release the paper. Please be patient. SegFix is a scheme that is complementary with the PointRend scheme, in fact, our method is related to the Signed-Distance-Transform from principle.

from openseg.pytorch.

My first question is actually this line. I am wondering why there are two segmentation outputs?

from openseg.pytorch.

@freaad This design follows the Auxilary Supervision Scheme (or DSN scheme) on the intermediate feature maps output from ResNet-101/Stage3, which is first adopted in PSPNet (Refer to Section 4 of PSPNet ) in the Semantic Segmentation task.

The first output x_dsn is mainly used to address the small gradient issue, in fact, during testing, we only use x as the final prediction.

from openseg.pytorch.

Thank you for your answer.

from openseg.pytorch.