Streaming 3d Reconstruction Eval

streaming-3d-reconstruction-eval

Geometric Context Transformer for Streaming 3D Reconstruction — Chen et al. (2026) (arXiv:2604.14141, 2026)

What this evaluates

Evaluates a model's ability to perform streaming camera pose estimation and 3D reconstruction over long video sequences. It probes long-range geometric consistency, drift resistance, and reconstruction fidelity across diverse indoor and outdoor environments.

Datasets

• Oxford Spires — total ?; splits: test (-1)
• ETH3D — total ?; splits: test (-1)
• 7-Scenes — total ?; splits: test (-1)
• Tanks and Temples — total ?; splits: test (-1)
• NRGBD — total ?; splits: test (-1)

Metrics

• ATE (primary) — range: meters
  • Absolute Trajectory Error, measuring the mean translation error between predicted and ground-truth camera poses after rigid alignment.
• AUC@15 — range: percent
  • Area Under the Curve of the cumulative pose error plot up to a 15-meter threshold.
• F1 (primary) — range: percent
  • F1 score for 3D reconstruction, computed as the harmonic mean of Accuracy and Completeness metrics.

Input / output format

Input: Causal sequence of RGB frames processed in a streaming fashion without access to future frames.

Output: Per-frame predicted camera poses (6-DoF) and/or a temporally consistent 3D point cloud.

Scoring recipe

def compute_ate(pred_poses, gt_poses):
    aligned = align_se3(pred_poses, gt_poses)
    return np.mean([np.linalg.norm(p - g) for p, g in zip(aligned, gt_poses)])

def compute_f1(pred_pts, gt_pts):
    acc = chamfer(pred_pts, gt_pts)
    comp = chamfer(gt_pts, pred_pts)
    return 2 * (acc * comp) / (acc + comp) * 100

Common pitfalls

• Streaming methods must not reset internal state between sequences for fair comparison.
• Oxford Spires has sparse (320 frames) and dense (3,840 frames) settings that test different aspects of drift.
• RPE-Rot is measured in degrees while ATE and RPE-trans are in meters.

Evidence (verbatim from paper)

We evaluate under two settings to test both aspects. In the sparse setting (320 frames, sampled every 12 frames), all method categories can run on our hardware, enabling a fair comparison across offline, optimization-based, and online approaches. As shown in Table 2, LingBot-Map achieves the best results on nearly all metrics. Despite operating in a streaming online manner, our method surpasses the strongest offline baselines by a large margin. LingBot-Map achieves an AUC@15 of 61.64, substantially exceeding the best offline method DA3 (49.84) and more than doubling VGGT (23.84). On trajectory-level accuracy, LingBot-Map reduces ATE from 12.87 (DA3) and 24.78 (VGGT) to 6.42.

Citation

@misc{chen2026geometric,
  title={Geometric Context Transformer for Streaming 3D Reconstruction},
  author={Chen et al. (2026)},
  year={2026},
  note={arXiv:2604.14141}
}

• arXiv: 2604.14141