the distortion and undistortion before and after generation

[Tcorpion]

I download some clips from nvidia/PhysicalAI-Autonomous-Vehicle-Cosmos-Drive-Dreams, and run Cosmos-Drive-Dreams/cosmos-drive-dreams-toolkits/render_from_rds_hq.py , I set the camera to ftheta, apply hdmap and lidar control, get a synthesized clip like bellow. After undistortion by Cosmos-Drive-Dreams/cosmos-drive-dreams-toolkits/rectify_ftheta_to_pinhole.py, there is still distortion like curved traffic signs.

ClipID="0079aad7-0fc5-4722-804d-e7c8c1b84263_570745200000_570765200000"

hdmap and lidar control:

{
"hdmap": {
"control_weight": 0.5,
"input_control": "placeholder",
"ckpt_path": "checkpoints/nvidia/Cosmos-Transfer1-7B-Sample-AV/hdmap_control.pt"
},
"lidar": {
"control_weight": 0.5,
"input_control": "placeholder",
"ckpt_path": "checkpoints/nvidia/Cosmos-Transfer1-7B-Sample-AV/lidar_control.pt"
}
}

hdmap control video:

0079aad7-0fc5-4722-804d-e7c8c1b84263_570745200000_570765200000_0.mp4

lidar control video:

0079aad7-0fc5-4722-804d-e7c8c1b84263_570745200000_570765200000_0.mp4

the caption is:

{
"Sunny": "The video shows a daytime highway scene from inside a vehicle. The road is brightly illuminated by the sun, with clear visibility of multiple lanes marked by white lines. Several vehicles ahead are visible, their brake lights glowing red. A large green road sign overhead indicates directions to various destinations such as Osnabr\u00fcck, Oldenburg, Bremen-Centrum, and Hanover. The surroundings are well-lit and open, with no pedestrians or other significant objects visible outside the vehicle. The environment appears calm, with minimal activity aside from the vehicles on the road.",
}

I follow the Cosmos-Drive-Dreams SDG Pipeline (scripts/generate_video_single_view.py) and get the bellow:

raw generation:

undistorted (rectify_ftheta_to_pinhole.py):

There are two questions:

1. Does the open sourced model only support producing raw data with distortion?
2. Is the distortion in the 2nd image above coming from the intrinsic parameter inaccuracy? or from the training by raw dash cam image with distortion?