search

Tsinghua-MARS-Lab / futr3d

Code for paper: FUTR3D: a unified sensor fusion framework for 3d detection
Apache License 2.0
268 stars 39 forks source link

the camera + radar detection result was different from the paper #45

Open YueWangTO opened 1 year ago

YueWangTO commented 1 year ago

![Uploading 16916505557965460992848583736149.jpg…]()

YueWangTO commented 1 year ago

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 commented 10 months ago

@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 commented 10 months ago

我们用的weight就是链接里的，你们遇到了什么问题？

yyunyu commented 7 months ago

@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。 image 1

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 commented 7 months ago

感谢您的来信，已收到