Jihan Yang
Jihan Yang
You can have a try. But as the cls head will change with more class, I am not sure how the performance will behave.
Have you 3D detection DA works such as ST3D? Also, as we know, ONCE benchmark already provide DA configs based on OpenPCDet, so what's the main issue?
In my toy case, it can evaluate directly if all config are set well, so we don't write any tutorial for evaluation. In you case, could you check why this...
We suppose the point clouds are saved in N*C as waymo, since different dataset may have different number of channels.
The feature is still downsampled to 1/8 due to the design of `BACKBONE_3D`. So if you would like to change the feature_map_stride, you also need to modify the downsampling schema...
Maybe you can refer to the evaluation tools about count those metrics. https://github.com/open-mmlab/OpenPCDet/blob/b61049fc6a7ec27442317bd04efc8e306d5e14b6/pcdet/datasets/kitti/kitti_object_eval_python/eval.py#L185
3D detector cannot overfit to a small subset without some specific modification. Just try to validate on the whole training set.
This might be caused by origin Pointpillar doesn't use gt_sampling (cannot remember clearly).
https://github.com/open-mmlab/OpenPCDet/blob/bd96d39af2389478820b34db7ad0272e9bb205db/pcdet/datasets/waymo/waymo_dataset.py#L436-L439
It seems that you haven't present all error messages.