We present MetricTrail, a streaming system for accurate camera motion estimation and dense reconstruction from long monocular videos. To keep inference efficient over long sequences, we revisit DPVO's sparse local estimator. We find that single-image metric depth substantially improves its trajectory accuracy, without retraining the correspondence network or enlarging the optimization window. Metric predictions initialize patch depths and remain as soft constraints during optimization. We further improve correspondence estimation through learned adapters that augment matching features and recurrent context with frozen DINOv3 representations. The correspondence estimator and adapters are trained on diverse short clips without additional long-sequence training. This modular design also allows the depth provider to be independently upgraded and supports supplied camera intrinsics and calibrated stereo observations without retraining. Under the reported RGB-only protocols, MetricTrail achieves the lowest average trajectory error among evaluated methods on KITTI, Oxford Spires, and VBR. Pairing the estimated poses with single-image depth predictions also produces strong dense reconstruction on indoor and outdoor benchmarks and coherent reconstructions of in-the-wild videos. With supplied intrinsics, the system processes KITTI-02 at 12.70 FPS with 3.21 GB peak GPU memory.
Figures & tables
Figure 2: System overview. For each input frame, single-image metric depth initializes patch depths and constrains BA, while frozen DINOv3 features augment matching and recurrent context through learned adapters. Within a bounded active window, patch correspondences drive joint pose and depth refinement, and the updated geometry guides the next matching step. The resulting poses align dense per-frame depth maps into a shared reconstruction.
Figure 3: Long-sequence camera trajectory estimation. Ground truth is gray and estimates are green . Our local estimator reduces deviations; loop closure further improves revisit alignment.
Method
00 4,541 fr. 3.7 km
01 1,101 fr. 2.5 km
02 4,661 fr. 5.1 km
03 801 fr. 0.6 km
04 271 fr. 0.4 km
05 2,761 fr. 2.2 km
06 1,101 fr. 1.2 km
07 1,101 fr. 0.7 km
08 4,071 fr. 3.2 km
09 1,591 fr. 1.7 km
10 1,201 fr. 0.9 km
Avg. ↓
Optimization / Offline
DPVO (Calibrated)
113.11
16.60
113.01
2.46
0.98
59.34
55.91
19.30
110.63
74.55
13.71
52.69
ViPE
25.87
148.34
53.35
5.29
1.95
24.22
11.72
1.03
18.51
20.55
4.71
28.68
LoGeR
54.98
36.57
36.20
4.27
1.62
33.41
11.78
13.33
22.92
17.89
8.06
21.91
LoGeR ∗
26.19
41.26
32.21
5.02
1.62
22.65
5.49
5.04
21.96
9.03
9.44
16.35
Scal3R w/ LC
4.30
45.29
42.06
3.36
1.74
3.30
2.49
2.03
36.69
12.32
6.46
14.55
Online
LongStream
92.55
46.01
134.70
3.81
1.95
84.69
23.12
14.93
62.07
85.61
21.48
51.90
Table 1: Full-sequence ATE (m) on KITTI Odometry. LC denotes loop closure; LoGeR ∗ uses alignment optimization. Underlining marks the best average in each baseline group and our default RGB-only result, which outperforms all baseline averages.
Offline/Chunk
Online
Ours
Dataset
VGGT- SLAM
MASt3R- SLAM
VGGT- Long
ViPE
LoGeR
LoGeR ∗
Infinite VGGT
Long Stream
CUT3R
TTT3R
Horizon Stream
LingBot- Map
w/o LC
w/ LC
Oxford Spires
23.42
12.12
7.41
7.06
4.32
4.28
31.76
15.92
14.74
9.40
8.75
5.15
1.77
1.77
VBR
78.20
92.64
91.23
46.20
28.02
25.24
102.90
77.93
66.25
64.99
25.30
27.53
16.36
10.88
Table 2: Average full-sequence ATE (m ↓ ) on Oxford Spires and VBR.
Figure 4: Qualitative dense reconstruction comparison. Our method produces more accurate reconstructions with finer point-cloud detail. The highlighted regions show clearer structural boundaries and more faithful local geometry, alongside detailed depth predictions.
Table 6
Components
ATE (m) ↓
Metric Depth
Multi- dataset
DINOv3
KITTI
Oxford
VBR
–
–
–
52.69
2.59
72.25
•
–
–
12.77
1.85
19.57
•
•
–
10.32
1.76
22.51
•
•
•
9.10
1.71
18.90
Table 5: Component ablation. We progressively add metric depth, multi-dataset training, and DINOv3 to DPVO and report mean calibrated ATE (m; ↓ ).
Figure 5: Qualitative ablation on KITTI 00. Metric-depth priors substantially reduce accumulated trajectory drift, while multi-dataset adaptation and DINOv3 further improve agreement between first-pass and revisit trajectories. On the same set of reference patches, the full estimator increases the fraction of correspondence endpoints within 5 px from 77.8% to 97.2% , providing a local diagnostic of the stronger visual evidence used by BA.
Dataset
Variant
Metric Prior
Pi3X
VDA
DA3Metric
Metric3Dv2
UniDepthV2
MoGe-2
KITTI
DPVO
27.92
15.16
14.05
11.62
11.18
12.77
+ Multi-dataset
24.90
8.97
10.03
10.57
10.80
10.32
+ DINO (Ours)
22.35
8.88
7.97
8.21
7.69
9.10
Oxford Spires
DPVO
5.57
2.25
1.28
1.42
3.89
1.85
+ Multi-dataset
4.53
1.87
1.32
1.33
6.04
1.76
Table 6: Metric-prior ablation. We report average ATE RMSE (m; lower is better). Each column uses the named metric prior for all three variants; multi-dataset training and DINO augmentation are added cumulatively. Best and second-best variants are highlighted within each metric-prior column and dataset.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: In-the-wild reconstruction across scale changes. Representative inputs span an indoor–outdoor–indoor transition. The complete reconstructions show whether the indoor and outdoor regions remain connected in a common scene.
Figure 7: Comparison with ViPE on in-the-wild sequences. Columns show the inputs, ViPE, ViPE with its metric prior replaced by MoGe-2, and MetricTrail . From top to bottom, the sequences contain a pronounced scale change, backward motion with a turn, and rapid forward motion. The shared-prior comparison tests whether replacing ViPE’s depth provider resolves the deformation.
Figure 8: Additional long-sequence reconstructions. KITTI examples appear above and VBR examples below. From left to right, we show a full-sequence reconstruction with selected camera locations, local point clouds, corresponding depth predictions, and additional reconstructions.
Method
ATE(m) ↓
Latency ↓
FPS ↑
Peak VRAM (GB) ↓
LingBot-MAP
82.29
0.2s
11.37
15.07
HorizonStream
84.62
1.2s
9.92
25.08
Ours (Uncalibrated)
24.69
175s
12.70
6.81
Ours (Calibrated)
18.83
0.7s
12.70
3.21
Appendix
Table 7: Runtime on KITTI-02 using an NVIDIA A100 GPU. Latency is time to first output; FPS is steady-state throughput.
Metric
LoGeR
LoGeR ∗
LingBot-MAP
HorizonStream
ViPE (w/ Global BA)
Ours (Uncalibrated)
Ours (Uncalibrated) w/ Global BA
ATE (cm) ↓
8.1
6.1
7.9
7.1
6.5
6.7
6.3
Appendix
Table 8: Full-sequence camera trajectory estimation on 7Scenes. We report mean ATE (cm; lower is better) over the 18 official test sequences, sampled at stride 5, after a proper per-sequence Sim(3) alignment. Global BA denotes global bundle adjustment.
Metric
DPVO (Calibrated)
CUT3R
MASt3R- SLAM
VGGT- SLAM
VGGT- Long
LoGeR
LoGeR ∗
LingBot-MAP
HorizonStream
ViPE (Uncalibrated)
ATE (cm) ↓
9.7
48.6
6.0
6.1
8.3
7.5
6.0
4.5
5.3
4.9
Appendix
Table 9: Full-sequence camera trajectory estimation on TUM RGB-D. We report ATE RMSE (cm; lower is better) over nine sequences and 8,694 frames processed at stride 1. Each sequence is aligned independently by a proper Sim(3) transform, and sequence-level errors are averaged equally. DPVO uses supplied intrinsics. Best and second-best results across both blocks are highlighted.
Method
R_01 2,898 fr.
R_04 5,253 fr.
R_06 7,808 fr.
R_07 8,010 fr.
R_08 12,328 fr.
Avg. ↓
LingBot-MAP
6.65
33.01
60.85
62.36
98.91
52.36
LingBot-MAP-Long
7.35
34.82
67.10
65.37
77.24
50.38
HorizonStream
7.19
35.63
54.55
42.70
90.63
46.14
LoGeR
4.23
10.87
22.44
23.01
91.11
30.33
LoGeR ∗
4.63
14.27
18.76
35.41
69.81
28.58
Ours (Uncalibrated)
3.09
4.85
15.97
10.14
9.57
8.73
Appendix
Table 10: Full-sequence ATE (m) on five sequences from LaMAria.
Uncalibrated
Calibrated
Metric
LingBot-MAP
ViPE †
HorizonStream
Ours
DROID-SLAM
Splat-SLAM
WildPose †
Ours
ATE (cm) ↓
3.48
2.59
21.58
2.70
16.17
8.71
0.39
1.65
Appendix
Table 11: Full-trajectory camera pose estimation on the dynamic Wild-SLAM MoCap benchmark. We report ATE RMSE (cm; lower is better), averaged equally over ten sequences. Methods are separated by whether camera intrinsics are provided at inference. † marks evaluated configurations that explicitly mask dynamic content. Best and second-best results are highlighted separately within each calibration regime. Our two variants use neither dynamic masks nor a dynamics-specific objective and operate without loop closure or global bundle adjustment.
Method
00 4,541 fr. 3.7 km
01 1,101 fr. 2.5 km
02 4,661 fr. 5.1 km
03 801 fr. 0.6 km
04 271 fr. 0.4 km
05 2,761 fr. 2.2 km
06 1,101 fr. 1.2 km
07 1,101 fr. 0.7 km
08 4,071 fr. 3.2 km
09 1,591 fr. 1.7 km
10 1,201 fr. 0.9 km
Avg. ↓
Optimization
VidMap † (Uncalibrated)
6.89
424.80
169.49
4.74
25.79
80.18
132.81
77.72
120.01
176.84
85.95
118.66
DROID-SLAM (Calibrated)
131.25
267.79
217.63
3.40
1.47
80.41
64.20
18.19
167.80
90.05
22.74
96.81
DPVO (Calibrated)
113.11
16.60
113.01
2.46
0.98
59.34
55.91
19.30
110.63
74.55
13.71
52.69
VidMap † (Calibrated)
7.50
234.88
35.74
2.00
14.66
1.66
49.00
13.50
123.27
8.97
2.87
44.91
DROID-W (Calibrated)
48.26
207.57
52.80
1.76
4.65
25.16
5.34
3.35
29.05
35.56
8.01
38.32
COLMAP
139.12
3.83
71.99
1.46
112.77
20.37
10.95
7.80
21.72
21.19
4.52
37.79
Appendix
Table 12: Full-sequence ATE (m) on KITTI Odometry. † marks method-native sparse output over keyframes. LoGeR ∗ requires alignment optimization. The strongest baseline average within each method group and our default RGB-only configuration are underlined.
Method
Keble-2 0.20 km
Keble-3 0.18 km
Keble-4 0.23 km
Keble-5 0.21 km
Observ.-1 0.27 km
Observ.-2 0.28 km
Christ-2 0.25 km
Christ-3 0.18 km
Christ-5 0.80 km
Bodleian-2 0.55 km
Avg. ↓
Offline/Chunk
VGGT-SLAM
20.56
16.20
24.71
30.02
3.97
7.42
30.09
4.90
28.13
68.24
23.42
MASt3R-SLAM
7.06
18.03
11.27
14.64
10.26
10.30
2.90
6.95
19.85
19.99
12.12
DA3-SLAM
2.86
3.88
4.34
7.97
18.69
3.87
8.89
6.93
15.48
44.78
11.77
VGGT-Long
9.07
9.37
11.77
3.96
6.03
4.05
2.41
0.48
14.38
12.59
7.41
ViPE
14.39
13.88
0.57
12.18
0.55
0.23
15.44
0.18
9.31
3.89
7.06
LoGeR
2.74
2.33
2.92
3.21
4.83
5.29
2.14
1.84
11.61
6.31
4.32
Appendix
Table 13: Full-sequence ATE (m) on Oxford Spires at stride one.
Method
colosseo_0 8,815 fr. 1.45 km
campus_0 12,042 fr. 2.73 km
campus_1 11,671 fr. 2.95 km
pincio_0 11,142 fr. 1.27 km
spagna_0 14,141 fr. 1.56 km
diag_0 10,021 fr. 1.02 km
ciampino_1 18,846 fr. 5.20 km
Avg. ↓
Offline/Chunk
MASt3R-SLAM
88.75
116.21
114.37
67.11
54.62
35.67
171.76
92.64
VGGT-Long
45.73
132.68
115.65
64.40
58.49
33.66
187.97
91.23
VGGT-SLAM
101.00
93.51
71.74
66.42
57.00
33.64
124.10
78.20
ViPE
79.22
46.07
63.10
15.42
20.33
35.75
63.48
46.20
LoGeR
31.77
27.90
30.80
17.96
21.33
32.25
34.16
28.02
LoGeR ∗
55.32
13.27
16.79
9.18
18.32
29.45
34.32
25.24
Appendix
Table 14: Full-sequence ATE (m) on VBR. The header reports the number of frames and route length of each sequence.
KITTI Odometry
Variants
00
01
02
03
04
05
06
07
08
09
10
Avg. ↓
Ours w/o Multi-dataset, DINO
7.61
76.22
17.17
2.54
0.79
8.02
5.44
2.83
10.38
4.80
4.64
12.77
Ours w/o DINO
6.05
30.52
21.36
2.20
3.67
4.92
10.19
3.83
17.84
7.98
5.00
10.32
Ours
5.22
33.21
18.83
2.71
0.59
5.90
6.18
3.13
11.94
8.39
3.95
9.10
Appendix
Table 15: Per-sequence component ablations: calibrated ATE RMSE (m) on KITTI Odometry, Oxford Spires, and VBR. All variants use monocular input, the same MoGe-2 prior, and no loop closure; Oxford Spires uses stride one. Only the lowest average within each dataset is bold.
Variant
colosseo_0
campus_0
campus_1
pincio_0
spagna_0
diag_0
ciampino_1
Avg. ↓
Ours w/o refinement
30.73
45.59
14.41
15.78
18.65
35.35
74.07
33.51
Ours
30.08
9.26
15.07
7.43
13.94
20.97
17.75
16.36
Ours w/o refinement w/ LC
23.44
37.27
9.49
14.46
13.36
34.80
62.85
27.95
Ours w/ LC
26.83
6.80
9.99
6.94
10.18
5.84
9.60
10.88
Appendix
Table 16: Effect of RGB-only focal refinement on VBR. We report full-sequence ATE (m; lower is better) with and without loop closure. The local estimator is held fixed.
Variant
colosseo_0
campus_0
campus_1
pincio_0
spagna_0
diag_0
ciampino_1
Macro abs. error
Ours w/o refinement
+7.68%
+8.26%
+1.25%
+8.04%
+7.18%
+8.78%
+8.37%
7.08%
Ours
+5.56%
+1.33%
+1.25%
-1.41%
+4.89%
-0.06%
+0.14%
2.09%
Appendix
Table 17: Retrospective focal analysis on VBR across ten independent runs. Entries are signed relative focal errors against sensor calibration, used only for reporting, never for routing, refinement, or model selection. Macro averages report absolute errors.
Variant
colosseo_0
campus_0
campus_1
pincio_0
spagna_0
diag_0
ciampino_1
Avg. ↓
Ours (Calibrated)
18.19
16.13
15.78
6.34
14.35
22.77
41.64
19.31
Ours (Uncalibrated)
28.38
8.46
15.42
7.73
13.42
21.70
18.77
16.27
Ours (Calibrated) w/ LC
12.92
13.06
7.46
5.50
12.68
8.34
25.82
12.25
Ours (Uncalibrated) w/ LC
21.54
6.02
9.75
7.18
9.55
5.53
13.06
10.38
Appendix
Table 18: Effect of camera-model choice on VBR. We report full-sequence ATE (m; lower is better) with and without loop closure. The local estimator is held fixed within each comparison.
Streaming 3D reconstruction incrementally builds geometrically consistent scene maps from incoming images or video under tight memory and latency limits.
Online monocular 3D reconstruction enables dense scene recovery from streaming video but remains fundamentally limited by the stability-adaptation dilemma: the reconstruction model must rapidly incorporate novel viewpoints while preserving previously accumulated scene structure. Existing streaming approaches rely on uniform or attention-based update mechanisms that often fail to account for abrupt viewpoint transitions, leading to trajectory drift and geometric inconsistencies over long sequences. We introduce PAS3R, a pose-adaptive streaming reconstruction framework that dynamically modulates state updates according to camera motion and scene structure. Our key insight is that frames contributing significant geometric novelty should exert stronger influence on the reconstruction state, while frames with minor viewpoint variation should prioritize preserving historical context. PAS3R operationalizes this principle through a motion-aware update mechanism that jointly leverages inter-frame pose variation and image frequency cues to estimate frame importance. To further stabilize long-horizon reconstruction, we introduce trajectory-consistent training objectives that incorporate relative pose constraints and acceleration regularization. A lightweight online stabilization module further suppresses high-frequency trajectory jitter and geometric artifacts without increasing memory consumption. Extensive experiments across multiple benchmarks demonstrate that PAS3R significantly improves trajectory accuracy, depth estimation, and point cloud reconstruction quality in long video sequences while maintaining competitive performance on shorter sequences.
Lanbo Xu, Liang Guo, Caigui Jiang +1
State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, China · University of East Anglia, Norwich, NR47TJ, United Kingdom
Recovering a dynamic 3D scene from a long monocular video is crucial for dense geometry, camera motion, and temporal correspondence to remain consistent in a shared coordinate system. Existing methods face two key challenges: (1) feed-forward reconstruction models provide accurate local predictions but are limited to short clips, and (2) long-range trackers preserve correspondences without producing dense sequence-level reconstruction. This paper presents LongDPM, a novel overlap-aware framework for scalable long-range monocular dynamic reconstruction. First, LongDPM processes long videos in overlapping chunks, keeping inference memory bounded by the chunk length. Second, it connects chunk-local coordinate systems through confidence-weighted registration with static-aware overlap abstraction. Third, it associates dynamic identities across chunk boundaries and fuses matched trajectories to recover coherent long-range 3D motion. Experimental results demonstrate that LongDPM achieves superior long-range reconstruction and tracking performance, reducing dense tracking EPE over V-DPM on PointOdyssey, Kubric-F, and Kubric-G, while obtaining the best TUM-dynamics ATE for camera pose estimation.
Scaling has powered recent advances in vision foundation models, yet extending this paradigm to metric depth estimation remains challenging due to heterogeneous sensor noise, camera-dependent biases, and metric ambiguity in noisy cross-source 3D data. We introduce Metric Anything, a simple and scalable pretraining framework that learns metric depth from noisy, diverse 3D sources without manually engineered prompts, camera-specific modeling, or task-specific architectures. Central to our approach is the Sparse Metric Prompt, created by randomly masking depth maps, which serves as a universal interface that decouples spatial reasoning from sensor and camera biases. Using about 20M image-depth pairs spanning reconstructed, captured, and rendered 3D data across 10000 camera models, we demonstrate-for the first time-a clear scaling trend in the metric depth track. The pretrained model excels at prompt-driven tasks such as depth completion, super-resolution and Radar-camera fusion, while its distilled prompt-free student achieves state-of-the-art results on monocular depth estimation, camera intrinsics recovery, single/multi-view metric 3D reconstruction, and VLA planning. We also show that using pretrained ViT of Metric Anything as a visual encoder significantly boosts Multimodal Large Language Model capabilities in spatial intelligence. These results show that metric depth estimation can benefit from the same scaling laws that drive modern foundation models, establishing a new path toward scalable and efficient real-world metric perception. We open-source MetricAnything at http://metric-anything.github.io/metric-anything-io/ to support community research.