search

tsunghan-wu / RandLA-Net-pytorch

:four_leaf_clover: Pytorch Implementation of RandLA-Net (https://arxiv.org/abs/1911.11236)
MIT License
122 stars 33 forks source link

About test. #7

Closed huixiancheng closed 3 years ago

huixiancheng commented 3 years ago

Thank you for the realization, great job. I was wondering if you submitted the seq11-21 prediction results to test codalab? As far as I know, there is a large difference between the valid and test of the Semantickitti dataset.

huixiancheng commented 3 years ago

Moreover would you like to show me where does the code do the RS(Random Sampling). As the author say in appendices: QQ截图20210705194841 https://github.com/tsunghan-mama/RandLA-Net-pytorch/blob/198f3cfff52f7ada9fc74baf73e1763c612fe043/dataset/semkitti_trainset.py#L39-L46 Look like is in here generate a key value of center_point and then use query to get number of cfg.num_points points. It's this mean RS(Random Sampling).?

tsunghan-wu commented 3 years ago
  1. For question 1: Due to the long inference time, we do not evaluate our reimplementation models on seq 11-21. However, we appreciate your comment and admit that there might be a large performance gap between val & test set.
  2. For question 2: Yes.
huixiancheng commented 3 years ago

Thanks a lot for your answer. Moreover, I want to raise your attention about this issues from offical repo? Have you check this part. Look like will a littile infuence the result. Moreover, look like the way you caculate the loss is not similar to offical repo. This is the log from Hu . log_train_SemanticKITTI_08.txt And this is your repo loss log which I gain. log_train.txt Look like a little different.

fenfenglitech commented 2 years ago

hi@huixiancheng,if you inference on semantickitti test area,i got a process killed problem when i test on sequences 13 ,19,21.could you give me some seggestions?