Bug in sample_points function in data_processor.py when training PointRCNN_iou

In the following code:

       if (num_points < len(points)): 
            pts_depth = np.linalg.norm(points[:, 0:3], axis=1)
            pts_near_flag = pts_depth < 40.0
            far_idxs_choice = np.where(pts_near_flag==0)[0]
            near_idxs = np.where(pts_near_flag ==1)[0]
            near_idxs_choice = np.random.choice(near_idxs, num_points - len(far_idxs_choice), replace=False)
            choice = np.concatenate((near_idxs_choice, far_idxs_choice), axis=0 ) \
                if(len(far_idxs_choice) >0) else near_idxs_choice
            np.random.shuffle(choice)

I encounter this error when training point_rcnn_iou:

Near_idxs_choice = np.random.choice (near_idxs, num_points - len(far_idxs_choice), replace = False) 
 File "mtrand.pyx", line 951, in numpy.random.mtrand.RandomState.choice
ueError: negative dimensions are not allowed

This should be possibly because the length of far_idxs_choice is larger than the default number of points16384.

[sshaoshuai]

Yes, it is just a hyper-parameter based on the point cloud range. Actually I am not sure whether the performance will drop if we simply simple 16384 points from the whole point cloud. This parameter aimed to balance the number of points in nearby and far away areas since the random sampling may sample much more points in the nearby area.

[sshaoshuai]

Thank you for the report! It is indeed a bug and I didn’t consider that situation since it is impossible in the KITTI dataset. You could submit a PR to fix it or I will find time to fix it.