Batch training actually mix points together?

[symphonylyh]

By looking at the code, it seems when batch_size > 1, point cloud coords and features from different scenes are stacked together. Namely, the data in each batch is multiple scenes overlapped altogether. Although there is a batch_idx at the 1st column coords[:,0], this information is not used in the backbone spconv.

Then by applying spconv, wouldn't it be computing features over an overlapped point cloud? I am assuming the spconv should capture local features in the point cloud, but this doesn't make sense if we apply to an overlapped point cloud. Can anyone check if my understanding is correct? Thanks

[saladcat]

same question