some question about bbox_head output

[muzi2045]

trying to export to onnx from prtrained waymo model from CenterPoint codebase (only detection vehicle class):

here is the original pointpillars bbox header(9 classes) export from OpenPCDet:

and the nuscenens bbox header export from CenterPoint (10 classes):

is there really need to separate the box_head into so many heads with [x,y],[z],[lwh],[yaw_sin, yaw_cos], [class] ? if here is like 10 classes to be trained, there will be 5*10 heads in bbox_heads with the same feature map, it looks like not a good way to deploy it in a real scene. @tianweiy

[tianweiy]

I agree. I guess combining all regression targets together won't hurt too much (maybe ~~0.5 mAP). I did it this way for easier coding. I will find time to support only two heads (class + all reg) and retrain the models but it will take me a while (probably after this paper gets in somewhere)

[tianweiy]

here is the original pointpillars bbox header(9 classes) export from OpenPCDet:

IS this openpcdet's KITTI PP? I think the recent nuScenes pp still follow the same setup (separate regression into 4 heads)

[muzi2045]

Yes, the original kiiti pp config not using multihead.

[tianweiy]

Okay, stay tuned. I will get some updates about this later when I get time and GPUs ... I am now mainly exploring the Waymo dataset

[muzi2045]

is there any Waymo dataset download mirror link in China? I can't download whole waymo datatset because it's so huge.... @tianweiy

[tianweiy]

I am not aware of any and the term disallows redistribution so you may need to ask the team directly...

[YznMur]

@muzi2045 , Could you please provide more information how to display a graph of the model using onnx ? how to provide an input here ?

[Qjizhi]

trying to export to onnx from prtrained waymo model from CenterPoint codebase (only detection vehicle class):

here is the original pointpillars bbox header(9 classes) export from OpenPCDet:

and the nuscenens bbox header export from CenterPoint (10 classes):

is there really need to separate the box_head into so many heads with [x,y],[z],[lwh],[yaw_sin, yaw_cos], [class] ? if here is like 10 classes to be trained, there will be 5*10 heads in bbox_heads with the same feature map, it looks like not a good way to deploy it in a real scene. @tianweiy

@muzi2045 , hi could you please provide the exporting onnx code? @tianweiy do you have any plan to support exporting onnx code?

[tianweiy]

I don't know any onnx so I don't have plan to support this in the near future

[jaydevshelat]

@muzi2045 @Qjizhi did you manage to do ONNX export? Any script/pointers you can give me will be very helpful. Thank You in advance!

[tianweiy]

there is one here https://github.com/CarkusL/CenterPoint but i never tried

[jaydevshelat]

Thank You @tianweiy ! I will look at it.