BOBrown
commented
1 year ago Hi! First of all, thank you for applying transfer learning and active learning to the detection task of point cloud data. This will be a very good approach and strategy. However, the projects you have showcased are only tested on a few publicly available benchmark datasets, yielding test results. Can you tell me how to use 3DTrans to train and test my own dataset? How can I utilize this project to further improve the detection capabilities of existing models such as CenterPoint and PV-RCNN++? Can you provide me with some practical methods, detailed steps, and suggestions that can be implemented? Thank you!
@luoxiaoliaolan Sorry for our delayed response due to recent numerous ddl.
A1: Followed by OpenPCDet, we will add the Custom Dataset usage in next version of 3DTrans, about at the end of July.
Q2: How to improve the CenterPoint and PV-RCNN++ A2: For cross-dataset usage for some typical detectors such as CenterPoint and PV-RCNN++, we observe two aspects that can effectively boost the cross-domain performance: 1) select a better source-domain model that can present a better generalizability compared with only using some off-the-shelf pre-trained model such as Waymo-pretrained model, where you can refer to our recent AD pre-training related work in arXiv AD-PT. 2) some customized methods to address domain differences, where we found that for point-cloud scene, the LiDAR beam and Object-size across different datasets are the main factors that hurt the model's cross-dataset application. Thus, maybe you can design some LiDAR beam down-sampling method. For example, we use the 16-beam Waymo data to perform the Waymo-to-nuScenes adaptation. Waymo-to-nuScenes.
Hi! First of all, thank you for applying transfer learning and active learning to the detection task of point cloud data. This will be a very good approach and strategy. However, the projects you have showcased are only tested on a few publicly available benchmark datasets, yielding test results. Can you tell me how to use 3DTrans to train and test my own dataset? How can I utilize this project to further improve the detection capabilities of existing models such as CenterPoint and PV-RCNN++? Can you provide me with some practical methods, detailed steps, and suggestions that can be implemented?
Thank you!