Datasets:

TianhangCheng7
/

IMU4DData

assets/chair.obj: the merged Blender chair geometry, in metres, Z-up, recentered at its local bounding-box centre. Geometry and size are unchanged.
assets/chair.json: shared asset ID, bounds/extent, mesh hash, source provenance, and the translation back into the original Blender mesh-local coordinates.
sessions/<session>.json: instance poses, quaternion, translation, local extent, source matrices, and provenance. All instances reference the same mesh file.
manifest.json: asset/session index.
sample_index.json: existing processed clips mapped to the session annotations.
qc/index.html: all ten sessions rendered back into their three cameras.
verification.json: numerical consistency checks.

Instances and placement sources

Sequences	Keys	Pose source
1–5	`chair`	Previously reviewed individual scene_1 fits
6–10	`chair` (near), `chair__2` (far)	User's unchanged `seq6.blend` placement, reused across the fixed setup

source_instance retains the earlier naming: near = chair_1 / Blender chair1, far = chair_0 / Blender chair2. The two instances have separate object IDs and the same asset_id: ncsa/meeting_room/chair and category: chair.

Seq 4–15 use the same physical camera setup according to the user. Every supplied calibration file anchors cam_01 with identity extrinsics and identical intrinsics; per-session refinement of the other cameras does not itself establish a changed cam_01 frame. Seq 6–10 therefore reuse the exact same two manual pose matrices, under the stationary-chair assumption checked against the background overlays. Seq 1–5 retain their own poses rather than assuming identical chair placement.

Coordinate convention

All saved object poses are in the calibration / cam_01 world, in metres: X = camera right, Y = camera down, Z = camera forward.

mesh_to_world uses column-vector convention:

world_vertices = local_vertices @ matrix[:3, :3].T + matrix[:3, 3]

The mesh is bounding-box centred. Consequently, translation_xyz is the world position of its bounding-box centre, not the original Blender object origin. source_mesh_to_world separately preserves that original pose. Recentering was checked to leave all placed world vertices unchanged (tolerance 1e-7 m).

track is a static length-10 vector: [qw, qx, qy, qz, tx, ty, tz, extent_x, extent_y, extent_z]. The extent is in the rotated object's local axes, not a world-aligned AABB. bbox_convention is local_extent_centered_at_track_translation. Repeat the vector over a clip's frame count when a [T,10] track is needed.

These files are standalone object annotations. Existing sample PKLs, WDS shards, human motion, and IMUs were not modified. In particular, do not concatenate these camera-world tracks with the packed IMU-world motion without an explicit coordinate conversion. sample_index.json provides sample-to-session bindings, not an implicit conversion or loader integration.

Reading an annotation

import json
from pathlib import Path
import numpy as np
import trimesh

root = Path('data/processed/ncsa/objects')
record = json.loads((root/'sessions/0915_scene_2_seq_6.json').read_text())
obj = record['objects']['chair']  # near chair
mesh = trimesh.load(root/obj['mesh_path'], force='mesh', process=False)
mesh.apply_transform(np.asarray(obj['mesh_to_world']))

Quality and reproducibility

The seq6 manual placement was accepted by the user after three-camera review. Its remaining mesh/seat/caster discrepancies are preserved; this release does not claim independently measured geometric ground truth. No scale fitting, deformation, or additional alignment was applied during publication. QC includes chair-to-chair occlusion but not occlusion by other room geometry.

From the IMU4D_dev root, reproduce into a new output directory:

OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 \
  python -m dataset_process.ncsa.export_chair_objects --out <new-output-directory>

The export refuses to overwrite an existing release. The source .blend, previous fit results and original chair mesh are retained.