Cityscapes 3D transfer into real world images about multi-task-transformer

[prismformore]

@jinfagang Hi there, thanks for showing interest in our project! You can definitely use inference.py to generate a 3D bounding box prediction for your particular case. However, please keep in mind that the model's accuracy may be significantly constrained as it has been trained on the Cityscapes dataset, which focuses on driving scenarios rather than everyday scenes. Furthermore, when visualizing the bounding box by projecting the 3D bounding box to 2D, having the right camera parameters is crucial, and these might not be easily obtainable for a generic internet image. Best of luck with your project! 😉

[lucasjinreal]

@prismformore Hello, suppose I have my own camera instrincis (for example, I using kitti or nscenes dataset), is the result can be gurranted?

[prismformore]

@jinfagang Yes I think the model can generate results for kitti and nusenes images using the correct camera parameters, but I have not tried that :) It will probably be much worse than models specifically trained on these datasets due to the difference in data distribution. And as the monocular 3D detection task is highly sensitive to depth estimation which is dataset-dependent, the results can be bad.

Caveat: I have not tried this before, so it can be wrong.

[lucasjinreal]

@prismformore thank u! your code seems has no license. consider using mit. ^.-

[prismformore]

@jinfagang Thanks for your suggestions! Yes we are considering changing the license to mit in the next update. I will update some codes for the Cityscapes-3D experiment soon. Stay tuned!