I have a couple of questions regarding training, inference and learning map from semantic-kitti.yaml. The number of classes are hard coded in cfgs/<MODEL>.yaml and semantickitti_utils.py, while the class probabilities (See 'content' in this config file) are not used for loss regularization? Did I understand this correctly?
For example, SalsaNext accepts a user specified semantic-kitti.yaml that contains learning map, class probabilities, color information etc during training. For inference, the same file has to be used for accurate results. Do you plan on supporting this? I am willing to help you with this, under the assumption that I understand your codebase correctly
Thank you for this great repo!
I have a couple of questions regarding training, inference and learning map from
semantic-kitti.yaml. The number of classes are hard coded incfgs/<MODEL>.yamlandsemantickitti_utils.py, while the class probabilities (See 'content' in this config file) are not used for loss regularization? Did I understand this correctly?For example, SalsaNext accepts a user specified
semantic-kitti.yamlthat contains learning map, class probabilities, color information etc during training. For inference, the same file has to be used for accurate results. Do you plan on supporting this? I am willing to help you with this, under the assumption that I understand your codebase correctlyUpdate: found the hardcoded weights