What makes the direction of boxes wrong

[xingzhezuomeng]

I don't know what makes the direction of boxes wrong.It happens in many pictures and scenes.Thank you

[xingzhezuomeng]

[#](

url)

[dsamvoel]

[#](

url)

Do you solve this problem?? I have same problem too

[xingzhezuomeng]

In branch 2.0(default),I don't know how to solve this problem. But I try in the branch main,and reconfigure the environment，this problem is solved.@dsamvoel

[dsamvoel]

Thank you for replying. I solved too.

[betterhalfwzm]

Thank you for replying. I solved too.

How did you solve it? Thank you

[maolin0815]

[#](

url)

Here is the reason：https://github.com/OpenDriveLab/UniAD/commit/275e8c94099300d1df6c34556e57b42cea84b235#diff-9abdef9bd472591ad787f42742420024eb055b7518ec60115e049d496dbeac06R31

the author has convert bbox_yaw and bbox_dims to mmdet3d v1.0.0rc6 format.

To get the correct box direction for visualization, just change: box_yaw = -box_yaw - np.pi / 2

[dsamvoel]

Thank you for replying. I solved too.

How did you solve it? Thank you

I just change the branch and reconfigure the conda environment. After that, the problem is solveld

[baiyeha]

[#]( url)

Here is the reason：275e8c9#diff-9abdef9bd472591ad787f42742420024eb055b7518ec60115e049d496dbeac06R31

the author has convert bbox_yaw and bbox_dims to mmdet3d v1.0.0rc6 format.

To get the correct box direction for visualization, just change: box_yaw = -box_yaw - np.pi / 2

Perhaps this is part of the problem. It seems that merely modifying the code here has not changed the visualization of the 3D box. There is still a problem with the direction of the detection box

[baiyeha]

[#]( url)

Here is the reason：275e8c9#diff-9abdef9bd472591ad787f42742420024eb055b7518ec60115e049d496dbeac06R31 the author has convert bbox_yaw and bbox_dims to mmdet3d v1.0.0rc6 format. To get the correct box direction for visualization, just change: box_yaw = -box_yaw - np.pi / 2

Perhaps this is part of the problem. It seems that merely modifying the code here has not changed the visualization of the 3D box. There is still a problem with the direction of the detection box

（1）I attempted to merely use self.pred_yaw in /tools/analysis_tools/visualize/utils.py in Uniad2.0 -> self.pred_yaw = -pred_yaw-np.pi/2 It seems that the orientation problem of the visual 3D inspection box can be solved.

（2）However, it's not quite clear whether such modifications made by the original author in version 2.0 will have an impact on the training results. Please offer your suggestions and ideas if you know about it.