search

lingorX / HieraSeg

CVPR2022 - Deep Hierarchical Semantic Segmentation - A structured, pixel-wise description of visual scenes in terms of the class hierarchy.
262 stars 24 forks source link

Reproductible results #7

Open simonwebertum opened 1 year ago

simonwebertum commented 1 year ago

Hi, very impressed by your work! I don't find exactly the same IoU when I run your configs files for training (for CityScapes, around 80% versus 81%; for LIP, around 55% versus 58%). I was wondering whether I missed something (like for example running x times and consider only the best run)? Thanks a lot!

lingorX commented 11 months ago

Can you provide the training log and system config for further analysis?

simonwebertum commented 10 months ago

Thanks for your reply!! I followed your config file https://github.com/lingorX/HieraSeg/blob/main/Pytorch/configs/deeplabv3plus/deeplabv3plus_r101-d8_512x1024_80k_cityscapes_hiera_triplet.py, with batch size of 8, and I enabled loss_hiera_focal. The training was done on Cityscapes Fine annotations as mentioned in your paper. Did you add coarse annotations for the training? It seems that the value of DeepLabV3+ you mention in your table is linked to Fine + Coarse annotations training https://www.cityscapes-dataset.com/benchmarks/#pixel-level-results