search

dragonfly606 / MonoCD

[CVPR 2024] MonoCD: Monocular 3D Object Detection with Complementary Depths
MIT License
24 stars 7 forks source link

Experiment using Nuscenes Dataset #5

Closed jesslyn1999 closed 3 months ago

jesslyn1999 commented 3 months ago

Hi, first of all, thank you for the great work and findings, mainly in using global clues to more accurately predict depth. I read the paper, specifically on Table.5, it points out that the experiment towards the NuScenes Dataset is also conducted. It would be so helpful for further exploration if the team could add the codes that support the NuScenes dataset experiment.

Thanks!

Screenshot 2024-06-06 at 6 23 25 PM
jesslyn1999 commented 3 months ago

I have no clue of how to get the road plane equation from the Nuscenes Dataset. Hoping for more insight...

image
dragonfly606 commented 3 months ago

Hi @jesslyn1999 , thank you for your attention to our work. The main purpose of Table 5 is to demonstrate the generalization of our method. The evaluation results on the NuScenes dataset are obtained by the model we trained on the KITTI dataset, so we did not utilize the ground plane equation of the Nuscenes dataset. To get the ground plane equation for a dataset you can refer to https://github.com/kujason/avod/issues/19. Apart from this, a simpler alternative is to use the 3D coordinates of all objects in an image for fitting (when the number of objects is more than three) and use a fixed ground plane equation y=1.562 (when the number of objects is less than three).

jesslyn1999 commented 3 months ago

oh I got it, thank you so much ! @dragonfly606