3D_IoU has no value during the training time

[Manuel-Z]

Thanks for the great work! It's useful for our work, but we had a problem during the training.

The 3D_IoU is always 0.00, leading to a lower performance than paper.

The training process just follows the code here, with CUDA=11.3 and torch==1.12.0, and we found this may caused by the Polygon function in iou_loss.py.

Have you got any idea about this problem, we look forward to your reply, thank you!

[dragonfly606]

Could you provide more training details? such as what dataset was used

[ManZ1919]

Thanks for your reply. We just followed the config used in the default config as monocd.yaml with KITTI dataset, and the training processes normally. log_09-25 11:55:03.txt

[2024-09-25 11:55:07,050] monocd.trainer INFO: Start training [2024-09-25 11:55:19,212] monocd.trainer INFO: eta: 15:38:43 iter: 10 loss: 73.0622 2D_IoU: 0.0754 3D_IoU: 0.0000 depth_loss: 20.5264 compensated_depth_loss: 1.5848 keypoint_depth_loss: 12.6731 hm_loss: 3.5794 bbox_loss: 0.9254 dims_loss: 3.2605 orien_loss: 1.8150 horizon_hm_loss: 3.7656 offset_loss: 0.4400 trunc_offset_loss: 0.0000 corner_loss: 6.0932 keypoint_loss: 13.9379 weighted_avg_depth_loss: 4.4608 depth_MAE: 0.8294 comp_cen_MAE: 0.5753 comp_02_MAE: 0.6096 comp_13_MAE: 0.6100 center_MAE: 3.1103 02_MAE: 3.4582 13_MAE: 3.3069 lower_MAE: 0.3881 hard_MAE: 0.7067 soft_MAE: 1.1548 time: 1.2141 data: 0.1173 lr: 0.00030000

and finally: [2024-09-25 23:07:45,613] monocd.inference INFO: Car AP@0.70, 0.70, 0.70: bbox AP:96.2938, 87.5118, 80.2193 bev AP:26.4258, 20.3711, 18.2670 3d AP:17.2279, 13.3145, 11.0903 aos AP:96.09, 86.88, 79.13 Car AP@0.70, 0.50, 0.50: bbox AP:96.2938, 87.5118, 80.2193 bev AP:63.8181, 48.7316, 43.0057 3d AP:57.2157, 43.2018, 37.6626 aos AP:96.09, 86.88, 79.13