Egoxtreme Eval

egoxtreme-eval

EgoXtreme: A Dataset for Robust Object Pose Estimation in Egocentric Views under Extreme Conditions — Yoon et al. (2026) (arXiv:2603.25135, 2026)

What this evaluates

Evaluates the robustness of 6D object pose estimation models under extreme real-world visual conditions, including severe motion blur, dynamic lighting, and smoke. It also benchmarks temporal tracking strategies in highly dynamic egocentric scenarios to assess motion-aware inference capabilities.

Datasets

• EgoXtreme — total ?; splits: test (-1)

Metrics

• ADD(-S) recall (primary) — range: percent
  • Percentage of frames where the symmetry-aware average distance between predicted and ground-truth model points is below a specified threshold (e.g., 0.1d, 0.2d, 0.3d). Computed as the mean of binary indicators (1 if distance <= threshold, else 0) across the test set.
• MSSD — range: percent
  • Maximum Symmetry-Aware Surface Distance. The maximum distance between any point on the predicted model surface and the closest point on the ground-truth model surface, accounting for object symmetries.
• MSPD — range: percent
  • Maximum Symmetry-Aware Projection Distance. The maximum 2D projection distance between predicted and ground-truth model points in the image plane, accounting for symmetries.

Input / output format

Input: RGB image frames, optionally preprocessed (deblurred, dehazed, or light-enhanced). Ground-truth bounding boxes are provided for baseline evaluation. For tracking experiments, the previous frame's pose estimate is also provided as a temporal prior.

Output: 6D object pose (3D rotation matrix/quaternion and 3D translation vector) relative to the camera coordinate system.

Scoring recipe

def compute_add_s_recall(pred_pose, gt_pose, model_pts, threshold):
    pred_pts = transform(model_pts, pred_pose)
    gt_pts = transform(model_pts, gt_pose)
    # Symmetry-aware distance: min over all valid symmetry transforms
    dist = min(np.linalg.norm(pred_pts - apply_symmetry(gt_pts, s), axis=1).max() for s in symmetries)
    return 1.0 if dist <= threshold else 0.0

# Aggregate over test set
recall = sum(compute_add_s_recall(p, g, m, threshold) for p, g, m in dataset) / len(dataset)

Common pitfalls

• Using detected bounding boxes instead of ground-truth boxes conflates detection errors with pose estimation performance, which the authors explicitly avoid by using GT boxes.
• Applying standard image restoration (deblurring/dehazing) often degrades pose estimation due to noise artifacts in extreme conditions, contrary to typical preprocessing assumptions.
• Direct temporal pose propagation fails in high-motion scenarios due to large inter-frame displacement, requiring hybrid confidence-based strategies rather than simple frame-to-frame tracking.

Evidence (verbatim from paper)

For evaluation metrics, we employ the ADD(-S) [2, 46] recall, along with the standard BOP metrics [35]: Maximum Symmetry-Aware Surface Distance (MSSD) and Maximum Symmetry-Aware Projection Distance (MSPD). End-to-end evaluation results using CNOS [32] detections are provided in Appendix C1.

Citation

@misc{yoon2026egoxtreme,
  title={EgoXtreme: A Dataset for Robust Object Pose Estimation in Egocentric Views under Extreme Conditions},
  author={Yoon et al. (2026)},
  year={2026},
  note={arXiv:2603.25135}
}

• arXiv: 2603.25135