How to avoid the dark spots in the output ?

[arulpraveent]

@vkhalidov I'm trying to use segmentation map for a dataset that I'm creating but I see some random dark spots in the segmentation map that is produced by the model I'm using the default configurations the following images will be helpful.
Segmentation map

Input Image

[vkhalidov]

@arulpraveent black regions on hands correspond to segments estimated as background, errors in this case. Have you tried the new version of DensePose which is now a project within detectron2? It is based on PyTorch (not Caffe2) and provides better models in the model zoo. In particular, R50 and R101 models with DeepLab head (R_50_FPN_DL_s1x and R_101_FPN_DL_s1x respectively) give much better AP scores than legacy (Caffe2) models.

[arulpraveent]

@vkhalidov Thanks for the input.
I tried using the latest version of detectron and the segmentation output looked really good.
But the model I am using depends on old DensePose's output colour gradients. Since the new DensePose version's output colour gradients are different from the old one, my model is failing.
Would you know if there is an option to change the new DensePose colour gradients to match that of old? and also is there a way to visualise only the masks without overlaying it on the input image.