search

happinesslz / LION

[NeurIPS 2024] Official code of ”LION: Linear Group RNN for 3D Object Detection in Point Clouds“
https://happinesslz.github.io/projects/LION/
Apache License 2.0
123 stars 9 forks source link

NuScenes Performace #3

Closed ihaohe closed 3 months ago

ihaohe commented 3 months ago

Thanks for sharing your great work!. I reproduced the model in nuscenes using LION-Mamba cfgs but the final performance is little lower than yours. Is it reasonable?

8 * A100, using LION-Mamba cfgs without any modifications

image

AlmoonYsl commented 3 months ago

Thanks for sharing your great work!. I reproduced the model in nuscenes using LION-Mamba cfgs but the final performance is little lower than yours. Is it reasonable?

8 * A100, using LION-Mamba cfgs without any modifications

image

It is normal for nuScenes and the performance fluctuation is not very large. You can train LION-Mamba with 48 epochs for better and more stable results.

happinesslz commented 3 months ago

@ihaohe Thanks for your attention! I think your reproduced results are acceptable. Our performance is mAP: 68.0 and NDS:72.1. Therefore, the performance gap is only 0.3 mAP and 0.1 NDS. I have tried to train 48 epochs and obtain a more stable and better performance with mAP of 68.2 and NDS of 72.3. The corresponding performance is as follows: image

ihaohe commented 3 months ago

Got it. Thank you