po03087/ego6d_rag

Ego6D — 3D scene features + head 6D pose + body 3D pose

Derived from the Nymeria dataset, 12 indoor venues (Loc_18 … Loc_29). All three modalities live in a shared per-venue SLAM world frame, so they overlay directly.

License: research / non-commercial (Nymeria-derived). Not affiliated with the Nymeria authors.

1 — 3D scene voxel features · scene_voxels/, scene_voxels_pca/

Per-venue 3D DINOv2 semantic feature voxels (0.15 m), built point-based: a DINO feature is registered onto each real SLAM point (fisheye projection + z-buffer occlusion + closest-observation) and then averaged per voxel.

• scene_voxels/<venue>.npz — centers[M,3] (voxel = mean of its points), feat[M,768] fp16, count[M], obs_dist[M], R(=0.15).
• scene_voxels_pca/<venue>.npz + pca.npz — cross-venue PCA-64 reduction of feat.

2 — Head 6D pose sequences · head_pose_6d/, head_pose_6d_local/

20 Hz windows (128 frames ≈ 6.4 s) of head 6-DoF pose (position + quaternion xyzw), from the Aria head SLAM trajectory.

• head_pose_6d/<venue>.npz — world frame: pose[N,128,7] (xyz+quat), mask[N,128], xyz[N,3] (window mean position = localization target), loc, act, split (0 train / 1 val), scene_id, scene_names, is_labeled.
• head_pose_6d_local/<venue>.npz — segment-local (each window's start translated to the origin, heading kept): pose_local[N,128,7], origin_world[N,3] (add back to recover world), + same metadata. (translate-only; pose_local[...,:3] = world − origin_world.)

3 — 3D body pose sequences · body_pose_3d/

Full-body SMPL pose, per scene, as 22 world-frame body joints at ~20 Hz (SMPL-H forward-kinematics on Nymeria's SMPL params).

• body_pose_3d/<scene>.npz — joints[N,22,3] (world), t_timecode_us[N], SMPL params global_orient[N,3], body_pose[N,69], transl[N,3], betas[10], fps.
• 22 joints (SMPL order): 0 pelvis, 1 L_hip, 2 R_hip, 3 spine1, 4 L_knee, 5 R_knee, 6 spine2, 7 L_ankle, 8 R_ankle, 9 spine3, 10 L_foot, 11 R_foot, 12 neck, 13 L_collar, 14 R_collar, 15 head, 16 L_shoulder, 17 R_shoulder, 18 L_elbow, 19 R_elbow, 20 L_wrist, 21 R_wrist. Bones = SMPL kinematic tree.

Coordinate frames & time sync

• Space: voxels, head pose, and body joints are all in the same venue SLAM world frame (verified: SMPL head joint sits ~0.10 m from the Aria head-camera position).
• Time: head pose uses head DEVICE time; body pose uses Nymeria global TIMECODE. Convert per scene with the head VRS: tc_ns = provider.convert_from_device_time_to_timecode_ns(t_device_ns), then match t_timecode_us.

Venues (12): Loc_18 … Loc_29. Head pose spans 149 scenes; body pose 148 scenes with SMPL; scene voxels one primary scene per venue.