Scaling of completed KITTI object

[darrenjkt]

Hi,

Thanks for your work. I was trying to reproduce your work on the KITTI shape completion.

I retrained the pcn_emd on the shapenet_car dataset provided with the following arguments:

parser.add_argument('--batch_size', type=int, default=32)
  parser.add_argument('--num_input_points', type=int, default=3000)
  parser.add_argument('--num_gt_points', type=int, default=16384)
  parser.add_argument('--base_lr', type=float, default=0.0001)
  parser.add_argument('--lr_decay', action='store_true')
  parser.add_argument('--lr_decay_steps', type=int, default=50000)
  parser.add_argument('--lr_decay_rate', type=float, default=0.7)
  parser.add_argument('--lr_clip', type=float, default=1e-6)
  parser.add_argument('--max_step', type=int, default=300000)

When I visualize the completed KITTI object with the incomplete one however, I noticed that the completed object's scaling is a bit smaller. The completed pcds were obtained using your test_kitti.py code as is. The image below shows the completed points (grey) vs the incomplete KITTI (coloured). When using your pcn_emd_car pre-trained model, there is no scaling issue.

Do you know what may be the cause of this scaling problem?