知乎移步过来的，NDS差距大

[yukaizhou]

博主你好，我是在知乎给您留言的NDS差距较大的人，参数设置是和您的提供的完全一致的（比对了您的log日志）。麻烦您看下，期待您的回复。

[>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 6019/6019, 7.9 task/s, elapsed: 763s, ETA: 0s Formating bboxes of pts_bbox Start to convert detection format... [>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 6019/6019, 85.5 task/s, elapsed: 70s, ETA: 0s Results writes to /tmp/tmpvx8pyo3x/results/pts_bbox/results_nusc.json Evaluating bboxes of pts_bbox mAP: 0.3049
mATE: 0.6782 mASE: 0.7065 mAOE: 1.5475 mAVE: 0.9790 mAAE: 0.2633 NDS: 0.2898 Eval time: 80.9s

Per-class results: Object Class AP ATE ASE AOE AVE AAE car 0.514 0.530 0.751 1.579 1.109 0.241 truck 0.211 0.701 0.796 1.576 0.922 0.246 bus 0.302 0.741 0.856 1.561 2.341 0.361 trailer 0.122 0.977 0.834 1.537 0.442 0.038 construction_vehicle 0.062 0.885 0.671 1.576 0.104 0.383 pedestrian 0.333 0.744 0.332 1.529 0.878 0.685 motorcycle 0.253 0.740 0.795 1.589 1.691 0.132 bicycle 0.228 0.496 0.813 1.694 0.345 0.020 traffic_cone 0.514 0.509 0.326 nan nan nan barrier 0.510 0.459 0.892 1.286 nan nan {'pts_bbox_NuScenes/car_AP_dist_0.5': 0.1689, 'pts_bbox_NuScenes/car_AP_dist_1.0': 0.446, 'pts_bbox_NuScenes/car_AP_dist_2.0': 0.6632, 'pts_bbox_NuScenes/car_AP_dist_4.0': 0.7764, 'pts_bbox_NuScenes/car_trans_err': 0.53, 'pts_bbox_NuScenes/car_scale_err': 0.7509, 'pts_bbox_NuScenes/car_orient_err': 1.5789, 'pts_bbox_NuScenes/car_vel_err': 1.109, 'pts_bbox_NuScenes/car_attr_err': 0.2406, 'pts_bbox_NuScenes/mATE': 0.6782, 'pts_bbox_NuScenes/mASE': 0.7065, 'pts_bbox_NuScenes/mAOE': 1.5475, 'pts_bbox_NuScenes/mAVE': 0.979, 'pts_bbox_NuScenes/mAAE': 0.2633, 'pts_bbox_NuScenes/truck_AP_dist_0.5': 0.0159, 'pts_bbox_NuScenes/truck_AP_dist_1.0': 0.1198, 'pts_bbox_NuScenes/truck_AP_dist_2.0': 0.2919, 'pts_bbox_NuScenes/truck_AP_dist_4.0': 0.4154, 'pts_bbox_NuScenes/truck_trans_err': 0.7015, 'pts_bbox_NuScenes/truck_scale_err': 0.796, 'pts_bbox_NuScenes/truck_orient_err': 1.5765, 'pts_bbox_NuScenes/truck_vel_err': 0.9219, 'pts_bbox_NuScenes/truck_attr_err': 0.2463, 'pts_bbox_NuScenes/construction_vehicle_AP_dist_0.5': 0.0, 'pts_bbox_NuScenes/construction_vehicle_AP_dist_1.0': 0.0144, 'pts_bbox_NuScenes/construction_vehicle_AP_dist_2.0': 0.0882, 'pts_bbox_NuScenes/construction_vehicle_AP_dist_4.0': 0.1474, 'pts_bbox_NuScenes/construction_vehicle_trans_err': 0.8849, 'pts_bbox_NuScenes/construction_vehicle_scale_err': 0.6715, 'pts_bbox_NuScenes/construction_vehicle_orient_err': 1.5763, 'pts_bbox_NuScenes/construction_vehicle_vel_err': 0.1039, 'pts_bbox_NuScenes/construction_vehicle_attr_err': 0.3834, 'pts_bbox_NuScenes/bus_AP_dist_0.5': 0.0279, 'pts_bbox_NuScenes/bus_AP_dist_1.0': 0.1692, 'pts_bbox_NuScenes/bus_AP_dist_2.0': 0.4251, 'pts_bbox_NuScenes/bus_AP_dist_4.0': 0.5864, 'pts_bbox_NuScenes/bus_trans_err': 0.741, 'pts_bbox_NuScenes/bus_scale_err': 0.8558, 'pts_bbox_NuScenes/bus_orient_err': 1.561, 'pts_bbox_NuScenes/bus_vel_err': 2.3411, 'pts_bbox_NuScenes/bus_attr_err': 0.3611, 'pts_bbox_NuScenes/trailer_AP_dist_0.5': 0.0, 'pts_bbox_NuScenes/trailer_AP_dist_1.0': 0.0196, 'pts_bbox_NuScenes/trailer_AP_dist_2.0': 0.1543, 'pts_bbox_NuScenes/trailer_AP_dist_4.0': 0.316, 'pts_bbox_NuScenes/trailer_trans_err': 0.977, 'pts_bbox_NuScenes/trailer_scale_err': 0.8336, 'pts_bbox_NuScenes/trailer_orient_err': 1.5366, 'pts_bbox_NuScenes/trailer_vel_err': 0.4418, 'pts_bbox_NuScenes/trailer_attr_err': 0.0381, 'pts_bbox_NuScenes/barrier_AP_dist_0.5': 0.2141, 'pts_bbox_NuScenes/barrier_AP_dist_1.0': 0.5074, 'pts_bbox_NuScenes/barrier_AP_dist_2.0': 0.6329, 'pts_bbox_NuScenes/barrier_AP_dist_4.0': 0.6858, 'pts_bbox_NuScenes/barrier_trans_err': 0.459, 'pts_bbox_NuScenes/barrier_scale_err': 0.8917, 'pts_bbox_NuScenes/barrier_orient_err': 1.2859, 'pts_bbox_NuScenes/barrier_vel_err': nan, 'pts_bbox_NuScenes/barrier_attr_err': nan, 'pts_bbox_NuScenes/motorcycle_AP_dist_0.5': 0.0415, 'pts_bbox_NuScenes/motorcycle_AP_dist_1.0': 0.1708, 'pts_bbox_NuScenes/motorcycle_AP_dist_2.0': 0.3609, 'pts_bbox_NuScenes/motorcycle_AP_dist_4.0': 0.4373, 'pts_bbox_NuScenes/motorcycle_trans_err': 0.7399, 'pts_bbox_NuScenes/motorcycle_scale_err': 0.7946, 'pts_bbox_NuScenes/motorcycle_orient_err': 1.5891, 'pts_bbox_NuScenes/motorcycle_vel_err': 1.6908, 'pts_bbox_NuScenes/motorcycle_attr_err': 0.1317, 'pts_bbox_NuScenes/bicycle_AP_dist_0.5': 0.0957, 'pts_bbox_NuScenes/bicycle_AP_dist_1.0': 0.2073, 'pts_bbox_NuScenes/bicycle_AP_dist_2.0': 0.2825, 'pts_bbox_NuScenes/bicycle_AP_dist_4.0': 0.3272, 'pts_bbox_NuScenes/bicycle_trans_err': 0.4959, 'pts_bbox_NuScenes/bicycle_scale_err': 0.8127, 'pts_bbox_NuScenes/bicycle_orient_err': 1.6942, 'pts_bbox_NuScenes/bicycle_vel_err': 0.345, 'pts_bbox_NuScenes/bicycle_attr_err': 0.02, 'pts_bbox_NuScenes/pedestrian_AP_dist_0.5': 0.0578, 'pts_bbox_NuScenes/pedestrian_AP_dist_1.0': 0.2455, 'pts_bbox_NuScenes/pedestrian_AP_dist_2.0': 0.441, 'pts_bbox_NuScenes/pedestrian_AP_dist_4.0': 0.5893, 'pts_bbox_NuScenes/pedestrian_trans_err': 0.7444, 'pts_bbox_NuScenes/pedestrian_scale_err': 0.3324, 'pts_bbox_NuScenes/pedestrian_orient_err': 1.5289, 'pts_bbox_NuScenes/pedestrian_vel_err': 0.8784, 'pts_bbox_NuScenes/pedestrian_attr_err': 0.685, 'pts_bbox_NuScenes/traffic_cone_AP_dist_0.5': 0.1956, 'pts_bbox_NuScenes/traffic_cone_AP_dist_1.0': 0.4944, 'pts_bbox_NuScenes/traffic_cone_AP_dist_2.0': 0.6474, 'pts_bbox_NuScenes/traffic_cone_AP_dist_4.0': 0.717, 'pts_bbox_NuScenes/traffic_cone_trans_err': 0.5086, 'pts_bbox_NuScenes/traffic_cone_scale_err': 0.3257, 'pts_bbox_NuScenes/traffic_cone_orient_err': nan, 'pts_bbox_NuScenes/traffic_cone_vel_err': nan, 'pts_bbox_NuScenes/traffic_cone_attr_err': nan, 'pts_bbox_NuScenes/NDS': 0.28977237815128287, 'pts_bbox_NuScenes/mAP': 0.30493887472544595}

[HuangJunJie2017]

It seem that you have use another version of mmdet3d to process the dataset. You have to rebuild the nuscenes dataset with the BEVDet code base.

[yukaizhou]

非常感谢您的回复。我的数据确实不是依赖您代码里的mmdet3d的版本生成的。我将会实验依赖您的代码里的mmdet3d版本进行数据的重新生成。再次感谢。

[eezhang123]

It seem that you have use another version of mmdet3d to process the dataset. You have to rebuild the nuscenes dataset with the BEVDet code base. I have a same problem, I exactly follow your REAMDme start step, but the NDS only 0.36. Besides, I download your r50 pth ,the NDS 45.7 meets the REAMDme mAp:0.3196 mATE:0.6989 mASE:0.7097 mAOE:1.5772 mAVE:0.3710 mAAE:0.1953 NDS:0.3623

[HuangJunJie2017]

@eezhang123 the main precision loss is in mASE and mAOE, you should process the dataset with BEVDet code base instead of other version of mmdet3d