Evaluation with empty predictions

[wojtat]

Hello!

I tried to run the model evaluation script with the pretrained model weights and a custom dataset. When the model doesn't predict anything in a point cloud, it seems that the function do_eval complains. I got the following error:

Traceback (most recent call last):
  File "evaluate.py", line 390, in <module>
    main(args)
  File "evaluate.py", line 374, in main
    do_eval(format_results, val_dataset.data_infos, CLASSES, saved_path)
  File "evaluate.py", line 63, in do_eval
    iou2d_v = iou2d(torch.from_numpy(gt_bboxes2d).cuda(), torch.from_numpy(det_bboxes2d).cuda())
  File "PointPillars/utils/process.py", line 432, in iou2d
    bboxes_x1 = torch.maximum(bboxes1[:, 0][:, None], bboxes2[:, 0][None, :]) # (n, m)
IndexError: too many indices for tensor of dimension 1

The iou2d function doesn't seem to account for empty detections. Do you have any suggestions how to fix this or how it should behave?

Thanks in advance for any help!

[zhulf0804]

Hello @wojtat , good question.

Update the code https://github.com/zhulf0804/PointPillars/blob/b9948e73505c8d6bfa631ffdf76c7148e82c5942/evaluate.py#L59 to

iou2d_v = iou2d(torch.from_numpy(gt_bboxes2d).reshape(-1, 4).cuda(), torch.from_numpy(det_bboxes2d).reshape(-1, 4).cuda())

For other commands using iou2d(), .reshape(-1, 4) is also added as above.

Then have a try ?

[wojtat]

Hi, after reshaping some of these detection arrays, it runs without crashing.

Thanks for the help!