search

sshaoshuai / PointRCNN

PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud, CVPR 2019.
MIT License
1.71k stars 425 forks source link

KITTI dataset format #189

Open uranuscc opened 3 years ago

uranuscc commented 3 years ago

Hi,

I was able to run the pointrcnn with provided KITTI point clouds set, and have difficulty to prepare my own dataset to replace the KITTI since there is no descriptions found for KITTI dataset format.

I have no idea about the details of the input data files. what data information and data format is required for preparing point clouds input, any plain docs on this?

May the auther guide or anyone help? Thanks.

Best Regards, CC

pytholic commented 3 years ago

I have a similar issue. I have my custom point clouds but I cannot find any guidance on preparing my dataset for these point cloud archtiectures.

kristiangulakerkm commented 2 years ago

For the binary files themselves (files in velodyne); you need to write your pointcloud to bytes, I use an array of bytes all made by struct.pack('ffff',x,y,z,r). For calib, you can simply mock the file with having each transformation/vector be the identity matrix. Labels are a little trickier, but in short: you only require the dimensions of each bounding box, it's centroid, and the rotation of the box along the y axis (yaw).