pre-trained models

[vobecant]

Hi,

first of all, let me thank you for this great work. Can I please ask you when do you intend to share the pre-trained models?

Thank you very much in advance.

Best, Antonin.

[JeffWang987]

We are working on optimizing the code framework, and the pretrained model will be released in a few weeks.

[qihao-plus]

My local reproduce with 8x3090, epoch 24 using camera-based baseline with loss_norm=False. The result is slightly lower than paper. Is that normal or correct ? @JeffWang987

2023-03-15 13:33:30,574 - mmdet - INFO - SC Evaluation
2023-03-15 13:33:30,574 - mmdet - INFO - +-----------+-------+
|   class   |  IoU  |
+-----------+-------+
| non-empty | 0.162 |
+-----------+-------+
2023-03-15 13:33:30,575 - mmdet - INFO - SSC Evaluation
2023-03-15 13:33:30,575 - mmdet - INFO - +----------------------+-------+
|        class         |  IoU  |
+----------------------+-------+
|         free         | 0.917 |
|       barrier        |  0.09 |
|       bicycle        | 0.057 |
|         bus          | 0.108 |
|         car          | 0.118 |
| construction_vehicle | 0.053 |
|      motorcycle      | 0.076 |
|      pedestrian      | 0.071 |
|     traffic_cone     | 0.043 |
|       trailer        |  0.04 |
|        truck         | 0.094 |
|  driveable_surface   | 0.213 |
|      other_flat      | 0.147 |
|       sidewalk       | 0.134 |
|       terrain        | 0.128 |
|       manmade        | 0.062 |
|      vegetation      | 0.092 |
|         mean         | 0.095 |
+----------------------+-------+
2023-03-15 13:33:30,577 - mmdet - INFO - Exp name: CAM-R50_img1600_128x128x10.py
2023-03-15 13:33:30,577 - mmdet - INFO - Epoch(val) [24][753]   SC_non-empty: 0.1620, SSC_free: 0.9170, SSC_barrier: 0.0900, SSC_bicycle: 0.0570, SSC_bus: 0.1080, SSC_car: 0.1180, SSC_construction_vehicle: 0.0530, SSC_motorcycle: 0.0760, SSC_pedestrian: 0.0710, SSC_traffic_cone: 0.0430, SSC_trailer: 0.0400, SSC_truck: 0.0940, SSC_driveable_surface: 0.2130, SSC_other_flat: 0.1470, SSC_sidewalk: 0.1340, SSC_terrain: 0.1280, SSC_manmade: 0.0620, SSC_vegetation: 0.0920, SSC_mean: 0.0950

[JeffWang987]

Training with loss_norm=False degrades the mIoU by ~0.5, and your reproduced results are reasonable.

[JeffWang987]

Already uploaded, check here.