search

emecercelik / ssl-3d-detection

40 stars 7 forks source link

unexpected result of pointpillars trained on nuscenes #7

Open cracker-Li opened 1 year ago

cracker-Li commented 1 year ago

Greetings!

I tried to modify the parameters as following according to the previous issue： set 'max_out=False' when training flow set 'max_out=True' and 'samples per gpu' from 2 to 1 when training det and testing det

Considering the long training time of flow, I used the nuscenes flow checkpoint you provided 屏幕快照 2023-05-17 下午2 48 52 and obtained the following training results（det trained for 24 epochs）, and found that there is a gap between the results and those in the paper.

屏幕快照 2023-05-17 下午2 37 36

I would like to ask if it is due to the issue of modifying the parameters or other reasons.

Additionally, I would like to know if the alternative training method mentioned in the paper was used to train pointpilots on Nuscenes. The above results were obtained without using the alternative training method.

HenryJunW commented 1 year ago

Hello @cracker-Li I can’t find the nuscenes flow checkpoint, may I ask what’s the link for that? Or which commit of the repo? Thanks!

ShaneCan commented 1 year ago

您好，我找不到nuscenes流量检查点，请问它的链接是什么？或者回购的哪个提交？谢谢！

same, i cant find it as well, if you know how to find it, plz tell me