Datasets:

---

nvidia

/

PhysicalAI-SmartSpaces

like 32

Follow

NVIDIA 29k

I Think......
Actually, the yaw doesn't necessarily indicate the direction of motion or its actual movement direction. yaw just say the geometric orientation of the object for representing rotated 3d bbox
In shorts, (yaw=-30d) == (yaw=+330d) -> same heading

yaw:	Euler angle in radians about the y-axis of the object-centered coordinate system defining the box’s heading in the world coordinate system. (Pitch and roll are assumed zero.)

maybe?

I Think......
Actually, the yaw doesn't necessarily indicate the direction of motion or its actual movement direction. yaw just say the geometric orientation of the object for representing rotated 3d bbox
In shorts, (yaw=-30d) == (yaw=+330d) -> same heading

yaw:	Euler angle in radians about the y-axis of the object-centered coordinate system defining the box’s heading in the world coordinate system. (Pitch and roll are assumed zero.)

maybe?

yeah, that's my first thought when I read the official guide: the yaw is "orienting" (azimuth) the bounding box in world coordinate system.
However I found that it makes more sense when the bounding box is orthogonal to the coordinate system, meaning the width, length and height are the sizes of the bounding box in X, Y and Z axis in world coordinate system (not oriented).
If I rotate the bounding box with yaw:

the person if moving toward the direction, but the bounding box is expanding in another direction

and also different type of object has completely different behavior regarding the yaw...

The only way I managed to make sense on every type of object is to NOT rotate at all, and add 90 degrees on yaw only on Person...

...

I Think......
Actually, the yaw doesn't necessarily indicate the direction of motion or its actual movement direction. yaw just say the geometric orientation of the object for representing rotated 3d bbox
In shorts, (yaw=-30d) == (yaw=+330d) -> same heading

yaw:	Euler angle in radians about the y-axis of the object-centered coordinate system defining the box’s heading in the world coordinate system. (Pitch and roll are assumed zero.)

maybe?

yeah, that's my first thought when I read the official guide: the yaw is "orienting" (azimuth) the bounding box in world coordinate system.
However I found that it makes more sense when the bounding box is orthogonal to the coordinate system, meaning the width, length and height are the sizes of the bounding box in X, Y and Z axis in world coordinate system (not oriented).
If I rotate the bounding box with yaw:

the person if moving toward the direction, but the bounding box is expanding in another direction

and also different type of object has completely different behavior regarding the yaw...

The only way I managed to make sense on every type of object is to NOT rotate at all, and add 90 degrees on yaw only on Person...

...

I got it what your saying!
It turns out that the ground truth 3d bounding box scale (i.e., width, height, length) values shouldn’t be used directly as-is when applying the yaw rotation. From what I’ve tested and visualized:

👉 The yaw seems to have been calculated assuming the width is always the longer side between width and height.
This assumption breaks if you use the raw scale = [width, length, height] values directly without checking their relative size.

So even though the GT provides width/height, to correctly reconstruct the rotated bounding box, I had to dynamically reassign width and height, ensuring:

width = max(width, length)
length= min(width, length)

and then apply the yaw. This way, the orientation of the bounding box aligns much better with the actual movement or heading direction — especially for non-human objects like Transporters or Carter.

PLUS..

And this is not just us to correct it, but GT itself is wrong, so I think it could be a problem with the HOTA calculation of the official leaderboard.

I found out similar phenomenon in Readme.md gif too.

Dear @mango16384 @visionNoob ,

Thank you for the detailed observations and helpful visualizations — this is indeed a known annotation issue that we are actively addressing.

We’re currently pushing updated train/val annotations to Hugging Face to address the discrepancies. Please refer to the changelog on the Hugging Face repository for details once the update is live.

In light of this, the total submission limit of Track 1 for the 2025 AI City Challenge will be increased to help teams adjust and re-evaluate their models. We’ve confirmed that the ground truth of test set remains unaffected.

For future inquiries, please feel free to also email us at [email protected] to ensure we can respond promptly.

Thanks,
Thomas

@mango16384 ,
I was wrong, as you said, the width, length and height are orthogonal to the world coordinate system this means it shoudn't be rotated by the yaw
Current 3d bounding box is not for rotated bbox but orthogonal bbox.