search

valeoai / DADA

Depth-aware Domain Adaptation in Semantic Segmentation
https://arxiv.org/abs/1904.01886
Other
114 stars 18 forks source link

Is any info from cityscapes needed besides raw images? #7

Open shumphries01 opened 3 years ago

shumphries01 commented 3 years ago

I'm a bit new to the field but am confused as to why DADA needs the ground truth labels of the cityscapes dataset (I noticed that these are included in the 'gtFine' folder in my data directory)? Are these ever used in training? I was under the impression the point of UDA is to circumvent the need to include labels in the target domain. I may have missed something, however.

Additionally, why is it necessary to map the Synthia classes to cityscapes' IDs (done in synthia.py file)? Thanks for the help! I'm trying to test on my own dataset and am doing my best to figure out what I need.

shumphries01 commented 3 years ago

I just consulted the cityscapes data structure bit in the ReadMe which clarified my first question! (ground truth is not used in training) However, if I could get some clarification on my second question (mapping between Synthia classes and cityscapes' IDs) still that'd be great! Thank you.