GT trajectory projection to camera image looks incorrect for PhysicalAI-AV dataset
Hi, thank you for releasing this amazing wrok!
I am trying to visualize the GT future trajectory by projecting it onto the camera image, but the result looks incorrect (trajectory does not align with the road geometry in the image), even though the turning direction is roughly correct.
Here are the details of what I did:
Dataset & clip
Dataset: nvidia/PhysicalAI-Autonomous-Vehicles
clip_id: 9779bda5-60c3-4f09-8dad-8b7fca09b496
GT trajectory
I use ego_future_xyz and ego_future_xyz_world returned by load_physical_aiavdataset
According to the loader implementation:
ego_future_xyz is expressed in the ego-local frame at t0
ego_future_xyz_world is in the world frame
For visualization, I either:
Directly project ego_future_xyz onto the t0 camera frame, or
Transform ego_future_xyz_world from world → ego(t0) local using egomotion, then project
Camera frame
I use the camera frame decoded at t0 (the last frame returned by decode_images_from_timestamps)
Image resolution exactly matches intrinsics (1920×1080)
Intrinsics and extrinsics are loaded directly from the dataset calibration
Observed issue
The projected GT trajectory does not lie on the road surface in the image
The overall turning trend (e.g., left/right turn) looks correct, but the trajectory is systematically offset
Changing a constant ego offset (e.g., ±4–5 m in x) does not fix the issue
Ego reference test points (e.g., [10, 0, 0] in ego frame) project to reasonable image locations
Question
Is ego_future_xyz guaranteed to be expressed in the same ego reference frame used by camera extrinsics (i.e., the ego frame origin for sensor calibration)?
Is there any known offset between the ego reference used for trajectory labels (e.g., rear axle / vehicle center) and the ego frame used in SENSOR_EXTRINSICS?
Are there any known caveats when projecting GT trajectories onto raw camera images for this dataset?
If anyone has encountered a similar issue or has guidance on the correct reference frame alignment, I would really appreciate it.
