How to reproduce the result on RTX 3090 machine ?

[kaixinbear]

Hello, cfzd !
I trained monoground on the kitti validation dataset and found that I can't reproduce the result on paper.

Since my lab doesn't have 2080Ti and RTX 3090 doesn't support cuda10, I use PyTorch version: 1.11.0+cu113, and rebuild DCNV2 based on https://github.com/lbin/DCNv2.
Others are keep the same with your config.What can I do to get the similar result as yours ?
Waiting for your reply.

[cfzd]

@kaixinbear
It is a very common problem that the results of mono 3D detection are not stable since it is composed of many individual tasks. It is also the same for our baseline SMOKE and MonoFlex.

The simple solution is that you can run it multiple times and record the best results.

[wenjie-yuan]

@cfzd @kaixinbear
My replication metrics are similar to yours, using 2 GPUs, 8 * 2 batches, 3090, torch1.9_cuda11.1, and the original ymal configuration file.
The result still lags behind the paper. Is there anything that needs to be modified in the configuration parameters?