I wish to train deeplab v3+ model on the MS COCO dataset. Has anyone successfully done that before? What training parameters were used to train the model?
I looked up in the issues in the Github repository but couldn't find any relevant information.
Things that I have tried before -
I have created the semantic segmentation labels and trained the model on complete data set. It seemed to under fit the data except the person class. I assigned weights to all the classes. But that didn't work out well.
I reduced the data set and trained my model on this reduced dataset but it didn't perform that well.
I am not sure if i am missing something. Any help will be really appreciated.