search

megvii-research / Far3D

[AAAI2024] Far3D: Expanding the Horizon for Surround-view 3D Object Detection
Other
134 stars 10 forks source link

Why train for 6 epochs? #32

Open lisanu opened 1 month ago

lisanu commented 1 month ago

Hello,

I noticed in your paper and code that you trained Far3D and StreamPETR for 6 epochs before comparison. However, the StreamPETR paper mentions training their network for 60 epochs for the case of NuScenes. Do you believe Far3D and StreamPETR can achieve good convergence in just 6 epochs on the Argoverse dataset? Could you explain why you chose to train for only 6 epochs?

exiawsh commented 1 month ago

The number of images of argoverse2 is 5x that of nuscenes, and the number of training iter is similar.

lisanu commented 1 month ago

Thanks for the response and great work, the results look impressive!