all the points are predicted to be the same class
mai4567 opened this issue · comments
Hi, could you provide a more detailed train.log?
Hi, I checked the log,
- It seem that background should be treated as a ignore segment?
- Also to ensure that trainining augmentation is aligned with validation alignment, maybe add an centershift after crop is better;
- Some scale parameter for PointGroup should be adjust, but I am not an expert, just do more experiment.