The results on the kitti dataset are not good？

[910399614]

Hi, I followed the released config file to train VoxelNeXt on the KITTI, but the detection results don't seem to be good?

Here is the epoch with the best results trained on my machine.

2023-05-30 13:25:51,687 INFO *************** Performance of EPOCH 72 *****************
2023-05-30 13:25:51,688 INFO Generate label finished(sec_per_example: 0.0634 second).
2023-05-30 13:25:51,688 INFO recall_roi_0.3: 0.000000
2023-05-30 13:25:51,688 INFO recall_rcnn_0.3: 0.935357
2023-05-30 13:25:51,688 INFO recall_roi_0.5: 0.000000
2023-05-30 13:25:51,688 INFO recall_rcnn_0.5: 0.873562
2023-05-30 13:25:51,688 INFO recall_roi_0.7: 0.000000
2023-05-30 13:25:51,689 INFO recall_rcnn_0.7: 0.633500
2023-05-30 13:25:51,715 INFO Average predicted number of objects(3769 samples): 12.604
2023-05-30 13:26:09,365 INFO Car AP@0.70, 0.70, 0.70:
bbox AP:90.4678, 89.2926, 88.3760
bev AP:89.0743, 86.4952, 83.4077
3d AP:86.5793, 76.6896, 74.6050
aos AP:90.45, 89.11, 88.13
Car AP_R40@0.70, 0.70, 0.70:
bbox AP:95.5828, 91.9155, 90.6527
bev AP:91.6419, 87.5357, 85.0153
3d AP:87.9198, 77.8969, 75.1770
aos AP:95.56, 91.71, 90.39
Car AP@0.70, 0.50, 0.50:
bbox AP:90.4678, 89.2926, 88.3760
bev AP:94.7389, 89.6084, 89.1113
3d AP:94.6626, 89.5168, 88.9023
aos AP:90.45, 89.11, 88.13
Car AP_R40@0.70, 0.50, 0.50:
bbox AP:95.5828, 91.9155, 90.6527
bev AP:96.8324, 94.3452, 93.6283
3d AP:96.7766, 94.1481, 91.5978
aos AP:95.56, 91.71, 90.39
Pedestrian AP@0.50, 0.50, 0.50:
bbox AP:76.9811, 73.3218, 69.6398
bev AP:65.7048, 61.0455, 56.5700
3d AP:62.2981, 56.9214, 52.1742
aos AP:75.42, 71.20, 67.12
Pedestrian AP_R40@0.50, 0.50, 0.50:
bbox AP:77.9146, 73.9838, 70.5503
bev AP:66.2811, 60.9713, 55.7163
3d AP:61.5019, 56.5423, 50.7263
aos AP:76.24, 71.67, 67.70
Pedestrian AP@0.50, 0.25, 0.25:
bbox AP:76.9811, 73.3218, 69.6398
bev AP:82.2534, 79.8235, 76.3670
3d AP:82.1528, 79.5656, 76.0224
aos AP:75.42, 71.20, 67.12
Pedestrian AP_R40@0.50, 0.25, 0.25:
bbox AP:77.9146, 73.9838, 70.5503
bev AP:84.9853, 81.9345, 77.8001
3d AP:84.8836, 81.4974, 77.4063
aos AP:76.24, 71.67, 67.70
Cyclist AP@0.50, 0.50, 0.50:
bbox AP:87.4056, 73.9096, 71.1011
bev AP:84.3527, 69.5915, 65.2032
3d AP:81.7647, 66.3772, 62.1164
aos AP:87.32, 73.52, 70.74
Cyclist AP_R40@0.50, 0.50, 0.50:
bbox AP:89.9100, 75.6070, 71.9232
bev AP:86.1329, 70.2481, 66.1244
3d AP:82.1792, 65.8815, 61.9532
aos AP:89.81, 75.18, 71.51
Cyclist AP@0.50, 0.25, 0.25:
bbox AP:87.4056, 73.9096, 71.1011
bev AP:86.7529, 71.9742, 69.1377
3d AP:86.7529, 71.9742, 69.1377
aos AP:87.32, 73.52, 70.74
Cyclist AP_R40@0.50, 0.25, 0.25:
bbox AP:89.9100, 75.6070, 71.9232
bev AP:89.0928, 73.4686, 69.6789
3d AP:89.0928, 73.4686, 69.6789
aos AP:89.81, 75.18, 71.51

[yukang2017]

#16

This is the result I get from this config.

I just wrote this config for an example on KITTI, without any hyper-parameters tuning. I think some hyper-parameters can be adjusted to get better performance.