About learning_map in semantic-kitti.yaml file

[aprilliuwei]

Many thanks to your team for the open source code,I want to use my own dataset to train in lidar-bonnetal, and i add some tags in the dataset that are not included in Kitti dataset, but I am not very clear about learning Map, and I want to know how to get those values. I tried to use the information you provided/ remap semantic_ labels. Py script file, but it doesn't seem to output anything.Please help me, thank you very much.

[jbehley]

the learning_map converts the original SemanticKITTI format, e.g., 10, 20, etc., into the 0,1,2,3,... labels used for learning. (Therefore, there are multiple labels mapped to "other-vehicle", like busses, etc.) The inverse learning map maps it again to the original SemanticKITTI labels that are used for evaluation. For evaluation on SemanticKITTI, we only account for labels that are defined. I guess other labels (say 123) are just mapped to 0 and ignored, i.e., will most likely decrease the IoU in the evaluation if they are not anyway 0 in the test data.

[aprilliuwei]

Thank you very much for your reply. I have successfully completed the training, and the visual prediction effect is very good during the training process, but when I use the trained model to test a single frame point cloud, as shown in Figure 1, it looks very terrible. When I use the pre-trained model to test the same single-frame point cloud, as shown in Figure 2,the prediction is normal. so I don't know why the pre-trained model has good results, but the effect of the model trained by myself is so bad.Thanks,Wish you all the best.

[juliangaal]

What kind of visualizer did you use here @aprilliuwei?