How to train the code on our own dataset?

[rockywind]

Hi, thank you for sharing the great work.
I train the code on our own dataset, the result is odd.
I think I misunderstand the coordinate conversion.
The bellow is my result.

[yshen47]

Thanks for trying! Could you share first on how the rendering looks like on your side using the kitti360 sequence? It should have the pedestrian rendered properly. If that is correct, you only need to swap the canvas image. meanwhile, you might need to check the intrinsics of your image, it seems your image is distorted

[rockywind]

Thanks a lot!

[rockywind]

Hi,
I found the insert objects' sizes are inconsistent with the background.
The traffic cone is the same high as the car.
How can I set the traffic looks like normal?

[yshen47]

Nice, at least the object is in the right orientation. To configure the scale, shadow intensity, direction, you can dive into our configurations file, located in configs/ . You can try tweak those parameters in the yaml file to give different scale and transform.

[rockywind]

Thanks.
How can I add a car in the image?
I add this config, but it doesn't work.