You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In the nuScenes paper not only d they write about the dataset, but also an idea to use points from multiple sweeps in order to . For instance tracking a moving point cloud solves the orientation and front/back ambiguities. It appears this could potentially be learned end-to-end simply by providing those timestamps.
@willyd I agree with you. The usage of the timestamp and the intensity should be not contradictory. The author use points = np.concatenate(sweep_points_list, axis=0)[:, [0, 1, 2, 4]], but not points = np.concatenate(sweep_points_list, axis=0)[:, [0, 1, 2, 3]], which is confusing.
@cvnerds Thanks for your answer. The use of timestamps as features can be considered as training moving point clouds.
But I still have a concern such an operation will also cause the overfitting of this dataset in an unreasonable manner.
as #254 (comment)
Hi,
@traveller59 Thanks for making this code available.
Could you please confirm that the behavior of this line is intended?
It looks like your are using the timestamp in place of the intensity. Is that correct? I think it should be
The text was updated successfully, but these errors were encountered: