How to train model on custom dataset?
RocketFlash opened this issue · 3 comments
Hello, @zhixinwang !
Thank you for your great work! Is it possible to train frustum-convnet on a custom dataset? What and where should I change in code?
@RocketFlash Current our code is mainly designed for KITTI dataset. If you want to use it for a custom dataset, the best way is to prepare the data according to KITTI format. For different categories in custom dataset, you can modify the information in datasets/dataset_info.py. The most important thing is you should care more about on the different coordinates. Contributions are welcomed to support other datasets.
@zhixinwang Thank you for the answer! I have another question: is it possible to train the network without camera data ? I mean if I have only lidar scans and 3D bboxes ? How is it better then to generate frustums ?
Our method leverages mature 2D detectors in RGB images to provide reliable frustum proposals.
If you don't have camera data, our method may not have an advantage. However, there is a possible method to generate frustum without RGB images. Try to first generate frustum proposals from range-view point cloud data. I do not have related experiments about this. You can try by yourself and welcome to share your experiment results.