val IoU of SalsaNext - Githubissues

TiagoCortinhal / SalsaNext

Uncertainty-aware Semantic Segmentation of LiDAR Point Clouds for Autonomous Driving

MIT License

417 stars 102 forks source link

val IoU of SalsaNext #13

Closed EthanZhangYi closed 4 years ago

EthanZhangYi commented 4 years ago

@TiagoCortinhal , I train the SalsaNext and in the log folder three checkpoints are saved, SalsaNext, SalsaNext_train_best and SalsaNext_valid_best. In SalsaNext_valid_best, the best_val_iou and valid_iou are 0.6101. However, I test SalsaNext_valid_best using infer.py and evaluate_iou.py , the average iou is 0.590, which is different from 0.6101. Why are they different?

TiagoCortinhal commented 4 years ago

Hey @EthanZhangYi !

That is a very good question. We barely adapt the eval setup from rangenet++. Nevertheless, I will investigate and come back to you once I have a clear answer!

shinya7y commented 4 years ago

trainer.py: projected images' pixel-wise IoU infer.py and evaluate_iou.py: point-wise IoU The latter will be suitable for point clouds SS.

EthanZhangYi commented 4 years ago

@shinya7y Thanks~