zmliu1122/mobjphys

Dataset Summary

This dataset contains object-centric HDF5 trajectories derived from Kubric/MOVi-style synthetic videos. Each .h5 file corresponds to one visible foreground object instance in a source scene and stores a normalized 512-point trajectory, object metadata, and physical parameters.

The dataset is provided in five variants:

Variant	Source scenes	H5 object files	Frames per object	Object set	Format
MOVi-A_verts_h5	1,200	7,810	480	CLEVR primitives: cube, cylinder, sphere	Points + per-frame mesh vertices/faces
MOVi-B_verts_h5	1,200	7,712	480	KuBasic shapes	Points + per-frame mesh vertices/faces
MOVi-spheres_verts_h5	1,200	7,809	480	Spheres only	Points + per-frame mesh vertices/faces
movib-single-f481_verts_h5	11,000	11,000	481	Single KuBasic object per scene	Points + per-frame mesh vertices/faces
movib-dual-collision-f481_points_h5	11,000	22,000	481	Two KuBasic objects per scene	Points-only + object poses

All variants contain 512 particles or sampled points per object. The first four variants also store normalized per-frame collision/mesh vertices and faces. The dual-collision variant is a compact points-only format: it stores the 512 sampled points over time plus object position/orientation trajectories, but does not store full per-frame collision mesh vertices or faces.

Dataset Variants

MOVi-A_verts_h5

A multi-object dataset using CLEVR-style primitives. Source scenes use CLEVR camera/background settings and 3-10 objects per scene. Object sizes are sampled from the CLEVR size set: small=0.7, large=1.4.

MOVi-B_verts_h5

A multi-object dataset using the KuBasic object set: cube, cylinder, sphere, cone, torus, gear, torus_knot, sponge, spot, teapot, and suzanne. Source scenes use random camera positions, colored backgrounds, and 3-10 objects per scene. Object sizes are sampled uniformly from [0.7, 1.4].

MOVi-spheres_verts_h5

A multi-object spheres-only dataset. Source scenes use CLEVR-style camera/background settings and 3-10 sphere objects per scene. This variant is useful for controlled experiments where object shape variation is removed.

movib-single-f481_verts_h5

A single-object KuBasic dataset with 1 object per scene and 1,000 scenes per shape type. Each sequence contains 481 frames. Object sizes are sampled uniformly from [0.7, 1.4].

Note: this variant is normalized per single-object scene and therefore has larger normalized object and motion scales than the multi-object variants.

movib-dual-collision-f481_points_h5

A two-object KuBasic dataset with exactly 2 visible foreground objects per scene. The two object types are sampled randomly with replacement from the KuBasic shape set, so the two objects may have the same or different shape type.

Each generated scene is required to contain an object-object collision within the 481-frame simulation. Scenes without a valid collision are discarded during generation and regenerated. This variant is therefore useful for studying pairwise rigid-body interaction and collision dynamics.

This variant uses a compact HDF5 format. It stores:

• normalized 512-point object trajectories in x
• normalized and world-space object positions
• object quaternions/orientations
• velocities and angular velocities when available
• scene normalization coefficients
• object metadata and physical parameters

It does not store full per-frame collision mesh vertices or faces.

Data Format

Each .h5 file contains one object trajectory.

Important fields for all variants:

Field	Shape	Description
x	(T, 512, 3)	Normalized particle or sampled point positions
vol	(512,)	Per-particle volume
scene_scale	scalar	Per-scene normalization scale
norm_coeffs/center_xy	(2,)	XY center used for normalization
norm_coeffs/scale	scalar	Same as scene_scale
norm_coeffs/z_min	scalar	Z offset used after scaling
shape_type	scalar string	Object shape name
asset_id	scalar string	Source asset id
shape_type_id	scalar int	Integer shape id
object_size	scalar	Object size sampled during generation
floor_friction, floor_restitution	scalar	Floor physical parameters
object_friction, object_restitution	scalar	Object physical parameters
frame_rate	scalar	Simulation frame rate
dt	scalar	Simulation timestep, equal to 1 / frame_rate
frame_start, frame_end	scalar	Frame range stored in the H5 file

Additional fields in the full mesh variants:

Field	Shape	Description
vertices	(T, V, 3)	Normalized mesh vertices over time
faces	(F, 3)	Mesh triangle faces
drag_point	(1, 3)	Initial object center / drag point
drag_force	(1, 3)	Initial velocity scaled by the scene normalization factor
drag_mask	(1, 512)	Particle mask for the object
C, F	(T, 512, 3, 3)	Placeholder deformation-related tensors
mat_type, gravity, nu, E	scalar	Simple material/physics parameters

Additional fields in movib-dual-collision-f481_points_h5:

Field	Shape	Description
points_only	scalar bool	True for the compact points-only format
positions	(481, 3)	Normalized object positions
quaternions	(481, 4)	Object orientations as quaternions
world_positions	(481, 3)	Raw world-space object positions
world_quaternions	(481, 4)	Raw object orientations as quaternions
velocities	(481, 3)	Normalized linear velocities, when available
world_velocities	(481, 3)	Raw world-space linear velocities, when available
angular_velocities	(481, 3)	Raw angular velocities, when available
mesh_vertex_indices	(N,)	Collision mesh vertex indices included in the 512 sampled points
num_mesh_vertices_in_x	scalar	Number of collision mesh vertices included in x
num_mesh_vertices_total	scalar	Total number of vertices in the source collision mesh
mesh_vertices_truncated	scalar bool	Whether the source collision mesh had more vertices than could be included
point_source	(512,)	Per-point source label

Here T=480 for MOVi-A, MOVi-B, and MOVi-spheres; T=481 for movib-single-f481_verts_h5 and movib-dual-collision-f481_points_h5.

Normalization

Coordinates are normalized independently per source scene. The exporter computes a scene-level scale from the full particle trajectory extent:

• XY is scaled so the maximum horizontal extent fits approximately within width 2.
• Z is scaled so the vertical extent fits approximately within height 1.
• The final scale is min(xy_scale, z_scale).
• XY coordinates are centered by the scene XY center.
• Z coordinates are shifted so the minimum normalized Z is 0.

The same scale is applied to particle positions, mesh vertices when present, particle volumes, velocities, and drag_force when present.

Because normalization is per scene, datasets with different scene composition can have different normalized object and motion scales even if their raw physical velocities are similar.

Scale Statistics

The following scene_scale statistics were measured directly from the H5 files:

Variant	Count	Mean	Median	P05	P95	Min	Max
MOVi-A_verts_h5	1,200	0.167073	0.165056	0.122669	0.216365	0.084758	0.266406
MOVi-B_verts_h5	1,200	0.174121	0.172751	0.126427	0.223434	0.096783	0.308355
MOVi-spheres_verts_h5	1,200	0.169431	0.166390	0.122931	0.222233	0.079567	0.288713
movib-single-f481_verts_h5	11,000	0.247027	0.236509	0.159122	0.363575	0.119241	0.587365
movib-dual-collision-f481_points_h5	11,000	0.260053	0.257494	0.189245	0.328570	0.098439	0.391424

The raw initial velocity distributions are comparable across the MOVi-style variants. However, movib-single-f481_verts_h5 and movib-dual-collision-f481_points_h5 have larger normalized coordinates and motion scales than the original multi-object MOVi-A/B/spheres variants because their scenes contain fewer objects and are normalized from smaller scene extents.

File Naming

Files are named as:

<scene_id>_<object_index>.h5

Dual-Collision Archive Reassembly

The dual-collision H5 archive is stored on Hugging Face as split tar parts because the full tar is larger than the single-file upload limit for the non-Xet LFS path. Expected full archive:

movib-dual-collision-f481_points_h5.tar
size: 58,836,049,920 bytes
sha256: e48ce3952ef3b3e73b875104be17060c61bc2af93b12f8490d57a10e4668c36e

Download the parts and manifest:

huggingface-cli download zmliu1122/mobjphys \
  --repo-type dataset \
  --include "movib-dual-collision-f481_points_h5.tar.part-*" \
  --include "movib-dual-collision-f481_points_h5.tar.parts.json" \
  --local-dir .

Reassemble the tar archive:

cat movib-dual-collision-f481_points_h5.tar.part-* > movib-dual-collision-f481_points_h5.tar

Verify the checksum:

sha256sum movib-dual-collision-f481_points_h5.tar