yingfei liu
yingfei liu
Hi, You can refer to https://github.com/megvii-research/PETR/issues/28 !
Hi, There is no difference between multiheadattention and PETRMultiheadAttention. The PETRMultiheadAttention is written here for the convenience of optimizing cross attention later.
Hi, In the training process, we will randomly rotate extrinsic (https://github.com/megvii-research/PETR/blob/main/projects/configs/petr/petr_vovnet_gridmask_p4_800x320.py#L161) to enhance the generalization performance. Therefore, in theory, it can be used for different extrinsic. Different frustum will be...
Hi, The streampetr has opensource in https://github.com/exiawsh/StreamPETR. Since the version of mmdetection3d is different, the streampetr will not into this repo. The streampetr repo has support PETRv1 and Focal PETR...
Hi, Thank you for your concern. 3D positional embedding is not an independent module in the code. In fact, 2D features will be added with 3D PE in multi-head attention...
Hi, When use the MMDetection3D v1.0.0rc5, the reverse_angle must be False. https://github.com/megvii-research/PETR/blob/main/projects/configs/petr/petr_r50dcn_gridmask_p4.py#L139
> > Hi, When use the MMDetection3D v1.0.0rc5, the reverse_angle must be False. https://github.com/megvii-research/PETR/blob/main/projects/configs/petr/petr_r50dcn_gridmask_p4.py#L139 > > Can you please explain why? mmdet3d reconstructed the coordinate system at v1.0.0rc5, and the...
The default data processing of mmdet3d will only process lidar's sweep data. We further run [https://github.com/megvii-research/Petr/blob/main/tools/generate _ sweep _ pkl.py](https://github.com/megvii-research/PETR/blob/main/tools/generate_sweep_pkl.py) to save the camera's sweep data in the sweep list.
@pianogGG [https://github.com/megvii-research/Petr/blob/main/tools/generate _ sweep _ pkl.py](https://github.com/megvii-research/PETR/blob/main/tools/generate_sweep_pkl.py) will not delete other files. It maybe the other problem. If you want to run the temporal version, I suggest you run [this one](https://github.com/exiawsh/StreamPETR)....
Yes, the grad norm always nan or inf, but it doesn't affect performance. You can use **optimizer_config = dict(type='Fp16OptimizerHook', loss_scale="dynamic", grad_clip=dict(max_norm=35, norm_type=2))** to make it more stable. Note that the...