Description:

We downloaded the xperience-10m-sample episode and performed a detailed analysis of the four fisheye camera extrinsics stored in annotation.hdf5. We believe the extrinsics for cam0 and cam2 are swapped (i.e., the calibration for cam0 actually belongs to cam2 and vice versa).
Evidence
We recovered the per-camera T_c_b (body-to-camera) transforms using get_T_camera_body() from HOMIE-toolkit, then computed the optical axis direction (camera z-axis in body frame) and pairwise angular separations.

According to the stored extrinsics:
cam1 optical axis yaw ≈ 22° (body frame)
cam2 optical axis yaw ≈ -166° (body frame)
cam1–cam2 angular separation: 113.5° (nearly opposite directions)
However, visual inspection of the actual video frames clearly shows that cam1 and cam2 are both front-facing cameras with heavily overlapping fields of view — they both see the same table, coffee cups, and the wearer's hands from similar perspectives.

Meanwhile, cam0 and cam3 show clearly different, non-overlapping perspectives (room corner / kitchen area), consistent with being the rear-facing pair.
Expected vs. Actual
If we swap cam0 and cam2 extrinsics, the geometry becomes consistent with the video content:
cam1 and cam0-with-swapped-extrinsics would form a front-facing stereo pair (yaw ≈ ±30° in rig frame, 51° separation)
cam2-with-swapped-extrinsics and cam3 would form a rear-facing pair
Additional observations
We also computed a rig-frame coordinate system from the four camera centers (best-fit plane, max residual < 0.02 cm). In the rig frame, all four cameras have nearly identical downward tilt (30°) and near-zero roll (<1.5°), which is consistent with a well-manufactured head-ring device. The yaw pattern in rig frame (±30° front pair, ±120° rear pair) makes perfect physical sense — but only if cam0 and cam2 labels are swapped.

Reproduction

# Using HOMIE-toolkit
from utils.calibration_utils import load_calibration_from_annotation_hdf5, get_T_camera_body
import h5py, numpy as np, math

with h5py.File("annotation.hdf5", "r") as f:
    calib = load_calibration_from_annotation_hdf5(f)

for cam in ["cam0", "cam1", "cam2", "cam3"]:
    T_c_b = get_T_camera_body(calib, cam)
    T_b_c = np.linalg.inv(T_c_b)
    z = T_b_c[:3, 2]  # optical axis in body frame
    yaw = math.degrees(math.atan2(z[1], z[0]))
    print(f"{cam}: yaw = {yaw:.1f}°")

# Output:
# cam0: yaw = 131.3°   <-- should be cam2?
# cam1: yaw = 22.0°
# cam2: yaw = -165.7°  <-- should be cam0?
# cam3: yaw = -69.1°

Environment

Dataset: ropedia-ai/xperience-10m-sample
Toolkit: Ropedia/HOMIE-toolkit (commit 5605e95)

Question

Is this a known issue? Does it affect only the sample episode or the full dataset as well? If this is confirmed, a fix in the annotation pipeline would be very helpful for downstream users.
Please check attached image for a clear visualization.