Pointcept / PointTransformerV3

[CVPR'24 Oral] Official repository of Point Transformer V3 (PTv3)
MIT License
787 stars 46 forks source link

Question about voxel size #51

Closed pycoco closed 5 months ago

pycoco commented 5 months ago

Is it supported different voxel size for x,y,z in PTv3?

Gofinge commented 5 months ago

Yes, of course. Adjust grid size in both data transforms and model configs.

pycoco commented 5 months ago

Yes, of course. Adjust grid size in both data transforms and model configs.

thanks! And if i use zyx order for grid_coords, where do I need to change the code in the ptv3 pipeline? My training loss does not converge, could u give me some suggestions?

Gofinge commented 5 months ago

And if i use zyx order for grid_coords, where do I need to change the code in the ptv3 pipeline

Just modified the serialization encoding code as "z-trans" or "hilbert-trans".

My training loss does not converge

There is not too much difference among orders with different axis priorities. The main purpose of introducing difference orders is to enable patch interactive like shift-window.

pycoco commented 5 months ago

thanks again, actually i try to modify "z-trans" or "hilbert-trans" related code, but the loss problem did not be fixed. Maybe I need to double check。and really look forward to you releasing the detection-related code and config as soon as possible.

Gofinge commented 5 months ago

Noted. BTW: code and config for detection is almost aligned with FlatFormer and is not build upon Pointcept (It will take a lot of time to make Pointcept to fully support object detection).

pycoco commented 5 months ago

thanks! i will have a try.