Object poses
Per session, motions/objects.npz holds every object instance on the body grid (same t / frame_index as
smplx.npz and markers.npz, 120 Hz), in the world frame (canonical Z-up metres).
| key | shape | meaning |
|---|---|---|
names |
(n,) | instance: BigBox0, LittleBox3, Table0; vision instances <Object>#1, #2, ... |
object_type |
(n,) | object type; meshes and scans in objects/<Object>/ |
mesh |
(n,) | mesh path from the release root (objects/BigBox/bigbox.ply), empty without a mesh pose |
pose_frame |
(n,) | mesh: pose of the scanned mesh. rigid_body: Nokov rigid-body frame (no mesh, e.g. Table, or no marker calibration for that object copy) |
source |
(n,) | mocap (object markers) or vision (sessions without object markers) |
T_world_obj |
(n, T, 4, 4) | pose_frame -> world; NaN where not valid |
valid |
(n, T) | usable pose |
quality |
(n, T) | 1 mocap verified (>= 3 markers agree with the pose), 2 mocap unverified (< 3 markers seen, none contradicts), 3 vision tracked, 4 vision rest fit, 5 vision bridged (moved with the hand or previous pose kept) |
motion |
(n, T) | 0 unknown, 1 resting, 2 moving (centre speed over 0.1 s > 0.05 m/s, dilated 0.25 s) |
held_by |
(n, T) | 0 none, 1 left, 2 right, 3 both hands (carries detected from the hand markers) |
T_rb_obj |
(n, 4, 4) | mocap mesh instances: Nokov rigid-body frame -> mesh frame |
calib_rms_mm |
(n,) | marker-to-mesh calibration residual |
Mesh poses from mocap: each physical object copy has its marker positions in the mesh frame
(objects/<Object>/markers.json, fitted at rest, upright on the floor), matched to the session's marker layout.
T_world_obj = T_world_rb @ T_rb_obj. Copies whose marker layout has no reliable calibration ship in the
rigid-body frame only (pose_frame = rigid_body): position, orientation and motion are measured, but the mesh is
not aligned. In this release that includes the LittleBox1 marker layout (markers on a tilted plane, no unique
upright fit), BigBox2 (too few rest windows), all OpenBox and LowStep copies, and objects without a mesh
(Table, Desk, HighRack).
motions/object_markers.npz (mocap sessions): raw object marker trajectories in the world frame, markers_m
(n, T, k, 3) with NaN where unseen, marker_names, template_m (markers in the rigid-body frame) and consistent
(marker agrees with the pose). The vendor files are in raw/mocap/<Object>.{csv,trc} (millimetres, Y-up).
motions/objects.json: per instance source, calibration, valid / verified / moving percentages; absent lists
mocap rigid bodies with under 1 s of valid pose (defined in the capture software but the object was not in the
take), which are left out of the npz files; and how the
session's object list was decided (object markers, or for sessions without them the objects recorded in other
takes of the same primitive, with hand separation deciding BigBox vs LittleBox where those takes differ).
Vision poses (sessions without object markers) come from model-based tracking in the head front stereo pair, anchored at fitted rest poses, at 30 Hz interpolated onto the body grid. Treat quality 5 as a guess.