search

drprojects / superpoint_transformer

Official PyTorch implementation of Superpoint Transformer introduced in [ICCV'23] "Efficient 3D Semantic Segmentation with Superpoint Transformer" and SuperCluster introduced in [3DV'24 Oral] "Scalable 3D Panoptic Segmentation As Superpoint Graph Clustering"
MIT License
545 stars 71 forks source link

Computing the NodeSize before sampling and are we really updating it ? #75

Closed gardiens closed 5 months ago

gardiens commented 6 months ago

Hello, Thank you for your amazing v2, I checked the new commits and I saw that you inverted the transform NodeSize and SampleSubNodes on the on-the-fly transform , Unless I am right, the NodeSize compute the number of low-level node for each high-level segment, but we sample some nodes afterwards and this feature doesn't seem to be updated by the sampling , Shouldn't we either compute it after sampling or update it during the sampling? Sincerely, Pierrick

drprojects commented 6 months ago

Hi Pierrick,

You are correct, the node_size attribute is not updated after SampleSubNodes. In our case, we decided this was a desired behavior for several superpoint-weighted computations where we would prefer using the initial superpoint sizes (position normalization, aggregated metrics computation, etc). It is possible that the opposite order would yield similar results. Feel free to try it and let us know if you find any improvement. Best,

Damien

drprojects commented 5 months ago

@gardiens, may I close this issue or are you still investigating it ?

gardiens commented 5 months ago

Hello , you can close it I will reopen once I got significant result :)

drprojects commented 5 months ago

Perfect, thanks for your prompt reply. Looking forward to hearing the results of your investigation :wink: