search

traveller59 / second.pytorch

SECOND for KITTI/NuScenes object detection
MIT License
1.73k stars 722 forks source link

Cropping data incorrectly for custom dataset #152

Open Magnetar99 opened 5 years ago

Magnetar99 commented 5 years ago

Hi @traveller59 ,

Thanks for this wonderful implementation. I was trying to use your implementation for custom pointcloud data which has been formatted exactly similar to Kitti but during pre-processing, the data crop does not work properly as seen below. It provides me with only one side of the pointcloud whereas all the bounding boxes have been provided. Could you please guide as to what mistake I'm making here? pc_9

traveller59 commented 5 years ago

KITTI point cloud crop need correct transpose matrix and projection matrix. I don't recommend to use KITTI format data, write custom dataset class or just don't crop point cloud.

Magnetar99 commented 5 years ago

@traveller59 thanks for your inputs. I am using the projection matrix only as a placeholder as I don't want to use the images to lidar correspondences. I've transformed the data to KITTI format as I am using the custom dataset to augment the KITTI dataset, is there any other alternative to handle this? Thanks. Hey @traveller59, can you please help me on this?

asgq123 commented 5 years ago

@Magnetar99 ,Hello, I'm doing the same thing with you , my dataset is only point cloud ,How do you modify the evaluation code? wish your help ,thanks a lot

940023303 commented 5 years ago

@asgq123 @Magnetar99 hello，I want to train the model using apollo dataset , but the dataset just have the label and pointcloud. and I don't know how to modify the evaluation function. can you help me?