search

VinAIResearch / ISBNet

ISBNet: a 3D Point Cloud Instance Segmentation Network with Instance-aware Sampling and Box-aware Dynamic Convolution (CVPR 2023)
Apache License 2.0
115 stars 23 forks source link

How to train the checkpoint of my datasets? #42

Closed wowluna closed 10 months ago

wowluna commented 11 months ago

Hi authors! Thank you for releasing this amazing work! Now, I have my own dataset similar to S3dis, but not the same scene. How can I train my own point cloud dataset and get a checkpoint?

Best wishes！ luna

ngoductuanlhp commented 10 months ago

Thank you for your interest in our work. To adapt our method to a new dataset, you need to create a new Dataset module, inherited from our CustomDataset, it will define how to read your own point cloud (including rgb and xyz) as well as the semantic and instance labels, and define the train-val split. If you do not have the superpoint for your data, you could create a "dummy" superpoint as we do in the Replica dataset here. Then you need to create a new config file, defining the number of semantic and instance classes, and the model's hyper-parameter.

wowluna commented 10 months ago

Thank you for your reply!

Best.