search

TiagoCortinhal / SalsaNext

Uncertainty-aware Semantic Segmentation of LiDAR Point Clouds for Autonomous Driving
MIT License
417 stars 102 forks source link

Output Predictions and Input Point Clouds Mismatch #74

Closed eyegotthis closed 1 year ago

eyegotthis commented 1 year ago

Hi there,

Thanks a ton for the amazing work, really appreciate. I was inferring the point cloud segmentations results from your script infer.py. I wanted segmentations for the training data for my own research but it seems that the output predictions do not match in size with the input point clouds. Tried the same for validations but this time there wasn't any mismatch. May I know what the reason could be for this and how I could obtain predictions that are equal in size to the input point clouds? Thanks a lot and looking forward to your response :)

Best, Esas

TiagoCortinhal commented 1 year ago

Hello!

Are you using SemanticKitti or another dataset?

eyegotthis commented 1 year ago

Hi there! Thanks for the response! Im using SemanticKitti

TiagoCortinhal commented 1 year ago

Can you please share with me the error message?

Sent from my iPhone

On 12 Apr 2023, at 15:06, eyegotthis @.***> wrote:



Hi there! Thanks for the response! Im using SemanticKitti

— Reply to this email directly, view it on GitHubhttps://github.com/TiagoCortinhal/SalsaNext/issues/74#issuecomment-1505245470, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AG2EFN6JD4S7KL7HJYNIWSTXA2SGPANCNFSM6AAAAAAW2RBJUU. You are receiving this because you commented.Message ID: @.***>

eyegotthis commented 1 year ago

There was not any error message thrown. The output predictions just had less number of points compared to the input point clouds when i opened the converted .label files back into numpy

eyegotthis commented 1 year ago

I have visualized these output predictions from the semantickitti API and everything seems normal except that the far off points were sometimes missing.

TiagoCortinhal commented 1 year ago

Did you turn off the Data augmentation when you do infer.py on the train split? By default the training has DA which includes dropping points.

TiagoCortinhal commented 1 year ago

You can change that here: https://github.com/TiagoCortinhal/SalsaNext/blob/a02fad97d646d4c132266ab79fbaea3ecfc237ed/train/tasks/semantic/dataset/kitti/parser.py#L326

eyegotthis commented 1 year ago

Thank you very much! I think this was the reason, will try this out today and write back if everything has worked out.

Edit: Checked. Problem fixed, much obliged!