LittlePey
commented
2 years ago Hi, it is possible to train our SFD on waymo dataset. But before that, you need to train a depth completion network. And before that, you need to generate depth labels for depth completion.
Given a sequence of point clouds, you can register a set of consecutive frames (e.g. 15 before and 15 after the current frame) with ICP (Iterative Closest Point), and accumulate these frames to generate a dense depth map which can be used as the depth label of the current frame. You can refer to the KITTI depth completion label generation process [1,2]. By the way, because waymo dataset provides a track id for each foreground object, you can directly use gt-boxes to align the point clouds of foreground objects in consecutive frames, instead of ICP.
[1] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. The International Journal of Robotics Research, 32(11):1231–1237, 2013. [2] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In 2017 international conference on 3D Vision (3DV), pages 11–20. IEEE, 2017.
In addition, you need to generate gt database of pseudo clouds for gt-sampling data augmentation. Hoping I understood your question correctly
d33dler
Is it possible to train the model with the waymo/custom dataset provided that the used training config file & model config respect waymo format as in other pcdet waymo cfgs?