the camera + radar detection result was different from the paper

[YueWangTO]

[YueWangTO]

I did not modify any parameters in the configuration file, but when I used the model provided by the author to verify the detection performance, the results were worse than those in the paper, especially the NDS index

[yyunyu]

@xyaochen 你好，我们一直无法复现你们的结果，可以分享一下你们的weights么，我们对你们的实验结果有些疑问

[xyaochen]

我们用的weight就是链接里的，你们遇到了什么问题？

[yyunyu]

@xyaochen 你好作者，radar+cam 模式下 我因为没有找到detr3d.pth, 使用了这个连接里面的https://drive.google.com/file/d/1YWX-jIS6fxG5_JKUBNVcZtsPtShdjE4O/view 显示模型和state dict不匹配，我也不知道要怎么修改state dict。

请问怎么训练才可以使得模型只使用detr3d.pth训练到像你们ckpt一样的效果？谢谢！

没有改动别的设置，然后结果是下面这样的：
Evaluating bboxes of pts_bbox
mAP: 0.0778
mATE: 1.1193
mASE: 0.5423
mAOE: 1.1838
mAVE: 1.1655
mAAE: 0.4092
NDS: 0.1438
Eval time: 5.0s
Per-class results:
Object Class AP ATE ASE AOE AVE AAE
car 0.185 1.106 0.203 0.724 0.290 0.073
truck 0.095 0.975 0.299 1.206 0.317 0.070
bus 0.051 1.330 0.235 2.253 3.008 0.235
trailer 0.000 1.000 1.000 1.000 1.000 1.000
construction_vehicle 0.000 1.000 1.000 1.000 1.000 1.000
pedestrian 0.188 1.047 0.286 1.373 0.837 0.267
motorcycle 0.138 1.153 0.405 1.397 0.206 0.333
bicycle 0.005 1.254 0.591 0.702 2.664 0.295
traffic_cone 0.116 1.329 0.404 nan nan nan
barrier 0.000 1.000 1.000 1.000 nan nan

[YueWangTO]

感谢您的来信，已收到