Datasets:

---

Hanibel

/

MOSIV

like 0

Hi, thank you for releasing the MOSIV dataset. I am trying to use the MOSIV/2objs sequences for physics simulation, and I would like to clarify the physical meaning of several metadata fields.

My main question is whether the rendered images / reconstructed 3D point clouds are expected to be directly consistent with the metadata values:

"fps": 15,
"gravity": -9.8,
"dt": 0.00020833333333333332

1. Rendered / reconstructed point-cloud motion does not seem to match g = 9.8 directly

I analyzed scene MOSIV/2objs/05. Using the rendered / reconstructed 4DGS point clouds, I estimated object-wise centroids from the first few frames. I used fps = 15, as specified in metadata.json.

From the first three frames, the estimated horizontal centroid velocities are approximately:

object / label 0:
v_horizon ≈ [-0.347,  0.369], speed ≈ 0.507

object / label 1:
v_horizon ≈ [ 0.346, -0.347], speed ≈ 0.490

However, the corresponding metadata initial velocities have horizontal speeds around:

obj1 metadata horizontal speed ≈ 0.84
obj2 metadata horizontal speed ≈ 0.85

So the observed point-cloud centroid motion is only about 0.58-0.60 of the metadata velocity scale.

I also estimated the vertical acceleration from the first four frames using second differences / quadratic fitting of the object-wise point-cloud centroids. The estimated acceleration along the point-cloud gravity direction is roughly:

object / label 0: a_vertical ≈ -3.4
object / label 1: a_vertical ≈ -4.0

This is much smaller than -9.8.

I understand that visible/reconstructed surface centroids are not exact centers of mass, and the estimate may be affected by occlusion, opacity, deformation, and reconstruction artifacts. However, the discrepancy is large enough that I am not sure whether the rendered images / reconstructed point clouds are intended to be in the same physical scale and time units as the metadata.

Could you clarify:

1. Are the rendered images / 3D reconstructions expected to be directly consistent with fps = 15 and gravity = 9.8?
2. Should the metadata initial_velocity be usable directly to reproduce the visible object motion?
3. Is there any global spatial scale or time scale that should be applied when using the metadata for physical simulation?
4. Is dt = 0.00020833333333333332 the actual simulation substep corresponding to the rendered sequence?

2. Meaning of collide_loc

I also noticed that across the MOSIV/2objs metadata files I checked, collide_loc is always:

"collide_loc": [0, 0, 0.14]

and collide_time is also constant:

"collide_time": 0.29166666666666663

Because collide_loc is identical across scenes, it seems unlikely to be the actual object-object collision location. It looks more like a fixed boundary plane / reference height / collision plane.

Could you clarify what collide_loc represents?

For example, is it:

1. a ground or boundary plane height,
2. a target collision/reference plane,
3. the intended object-object collision location,
4. or something else used internally during data generation?

This matters for simulation because I need to know whether to treat [0, 0, 0.14] as a physical boundary constraint or as an event annotation.

3. Minor coordinate-frame clarification

One minor coordinate-related observation: camera centers in metadata.json and the camera transforms suggest a coordinate conversion of approximately:

metadata/world:  (x, y, z)
render/camera:   (x, z, -y)

For example, a metadata camera center like:

[-0.339, 2.85, 0.873]

appears in the transform as:

[-0.339, 0.873, -2.85]

Could you briefly clarify which coordinate frame is used by:

initial_location
initial_velocity
gravity
collide_loc
mpm_lower_bound / mpm_upper_bound

Thanks again for the dataset. Any clarification on the physical scale, gravity/time units, and collide_loc semantics would be very helpful for using MOSIV in physics-based reconstruction or differentiable simulation.