search

PRBonn / lidar-bonnetal

Semantic and Instance Segmentation of LiDAR point clouds for autonomous driving
http://semantic-kitti.org
MIT License
945 stars 205 forks source link

About train my dataset result #94

Closed xdtzzz closed 2 years ago

xdtzzz commented 2 years ago

Hi, thanks for your code! I come again...... I try to train my dataset, this is my result in /log file 2021-11-18 16-12-40 的屏幕截图 2021-11-18 16-13-54 的屏幕截图 It seems that the predictions pictures are a little correct. But when I change it to onnx model and try to run your RangeNet code, it shows very bad, so that every class isn't correct. The following is that I train my part of dataset and run RangeNet code. 2021-11-18 16-14-17 的屏幕截图 The following is also that I train my other dataset and run RangeNet code. 2021-11-18 16-21-12 的屏幕截图

These two result above of difference is the train, valid and test dataset is not totally same.

Why the result is so bag...... Could you give me some suggestions?

jbehley commented 2 years ago

Sorry for the late reply. Currently I have no clue what is wrong. Did you check that your modifications to the yaml file are also correct in terms of the label mapping? We use a mapping from the semantic kitti classes (e.g., 10 is car, 30 is pedestrian) to the classes (0, 1, 2, ...) that is used for training. For visualization we map the classes back to the original classes. Thus you not only have to modify the one way but also the other way.

xdtzzz commented 2 years ago

Thanks your reply, i close the issue