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
508 stars 65 forks source link

shapenet dataset #119

Closed meehirmhatrepy closed 1 month ago

meehirmhatrepy commented 1 month ago

what modifications would make this approach work with shapenet dataset where the goal is to segment parts of pointclouds like Aeroplane, Bike, etc. which are not necessarily aligned with axis. Any suggestions would be of great help. Thankyou.

drprojects commented 1 month ago

Our method is essentially built for dealing with very large scenes of varying sizes. ShapeNet, on the other hand, is an easier setup with isolated objects, without acquisition artifacts, and reduced number of points. Yet, Superpoint Transformer should definitely work on ShapeNet too. You would obviously need to adapt the preprocessing parametrization to adjust to the desired voxel resolution, partition coarsness, partition graph size, etc. It would definitely be an interesting to this project, feel free to send us a pull request if you implement a new ShapeNet dataset !

PS: I am currently working on a tutorial for using Superpoint Transformer on a custom dataset, which might provide some guidelines for what you are trying to do. It will be released for an live demo session on June 27 with @florentPoux, whom will likely send some info in the coming days, so stay tuned :wink:

meehirmhatrepy commented 1 month ago

Would you please guide me with few parameters to change please? I have a custom dataset of various types bottles in which i wish to do part segmentation like outer part, lid, base, etc. As there is no alignment in point clouds with respect to z axis, the predictions are very off. how do i avoid this? Like what if a point clouds are not aligned with respect to z axis ?

drprojects commented 1 month ago

If your dataset if purely made of bottles, you could easily align all your clouds along the Z axis by using PCA