We introduce OpenVO, a novel framework for Open-world Visual Odometry (VO) with temporal awareness under limited input conditions. OpenVO effectively estimates real-world-scale ego-motion from monocular dashcam footage with varying observation rates and uncalibrated cameras, enabling robust trajectory dataset construction from rare driving events recorded in dashcam. Existing VO methods are trained on fixed observation frequency (e.g., 10Hz or 12Hz), completely overlooking temporal dynamics information. Many prior methods also require calibrated cameras with known intrinsic parameters. Consequently, their performance degrades when (1) deployed under unseen observation frequencies or (2) applied to uncalibrated cameras. These significantly limit their generalizability to many downstream tasks, such as extracting trajectories from dashcam footage. To address these challenges, OpenVO (1) explicitly encodes temporal dynamics information within a two-frame pose regression framework and (2) leverages 3D geometric priors derived from foundation models. We validate our method on three major autonomous-driving benchmarks - KITTI, nuScenes, and Argoverse 2 - achieving more than 20 performance improvement over state-of-the-art approaches. Under varying observation rate settings, our method is significantly more robust, achieving 46%-92% lower errors across all metrics. These results demonstrate the versatility of OpenVO for real-world 3D reconstruction and diverse downstream applications.
Figures & tables
Figure 2: Overview of OpenVO . We propose a novel temporal-dynamics-informed, geometry-aware visual odometry system. Our method takes consecutive dashcam frames as input and extracts both temporal and geometric representations for robust egomotion estimation. The Time-Aware Flow Encoder (Sec. 3.1 ) leverages a Differentiable 2D-Guided 3D Flow module and time-conditioned embeddings to model motion dynamics across varying observation rates, while the Geometry-Aware Context Encoder (Sec. 3.2 ) incorporates metric depth and intrinsic priors to build a consistent 3D geometry structure of the scene. Finally, the World-Coordinate Egomotion Decoder (Sec. 3.3 ) predicts accurate world-coordinate egomotion trajectories from the fused dynamic-geometric representation.
Method
KITTI 00–10 (10Hz)
nuScenes (12Hz)
Argoverse 2 (10HZ)
terr
rerr
ATE
serr
terr
rerr
ATE
serr
terr
rerr
ATE
serr
TartanVO [ 53 ]
13.85
3.27
103.07
-
10.27
6.35
6.26
-
11.17
5.30
7.03
-
DPVO [ 48 ]
8.31
2.37
78.53
-
4.34
2.85
2.66
-
2.66
1.25
1.59
-
ZeroVO + [ 23 ]
6.81
2.69
104.69
0.06
9.74
4.37
6.03
0.12
4.64
2.83
3.05
0.09
XVO [ 22 ]
16.82
3.84
168.43
0.17
12.75
5.11
8.30
0.16
9.13
4.86
5.70
0.12
M+DS [ 17 , 47 ]
14.22
2.72
154.77
0.09
17.08
1.46
10.46
0.18
16.67
1.79
8.51
0.13
Table 1: Generalizability results. Comparison with existing methods on 3 large-scale driving benchmarks using standard metrics: translation ( terr ), rotation ( rerr ), absolute trajectory (ATE), and scale error ( serr ). OpenVO is trained exclusively on nuScenes and evaluated on KITTI and Argoverse 2 with unseen camera setups, as well as unseen regions of nuScenes, to assess cross-domain generalization. Shaded rows denote methods not directly comparable to OpenVO (either rely on ground-truth intrinsics or auxiliary modality). “–” indicates unreported values; + denotes training with additional text guidance and YouTube data, and ‡ suggests the use of the same Depth [ 17 ] and Intrinsic [ 68 ] priors from foundation models as ours. OpenVO has 2 variants: ✓ with differentiable and × with non-differentiable 2D-Guided 3D flow. The best and second best results are marked as bold and underline among intrinsic-free, vision-only models.
Figure 3: Qualitative results . We present trajectory prediction results on the KITTI and nuScenes datasets. Compared to ZeroVO ‡ , both variants of our method — differentiable ( OpenVO-diff ) and non-differentiable variants ( OpenVO-nodiff ) of our 2D-guided 3D flow — achieve higher trajectory prediction accuracy and consistency, surpassing the current state-of-the-art.
Setting
KITTI 00–10 (10Hz)
nuScenes (12Hz)
terr
rerr
ATE
serr
terr
rerr
ATE
serr
K=2
9.68
3.55
108.48
0.07
9.47
3.90
6.63
0.11
K=4
9.43
3.54
101.53
0.06
9.95
3.60
6.74
0.12
K=8
8.97
3.43
93.23
0.05
9.04
3.86
5.91
0.10
K=16
9.18
3.72
95.11
0.06
9.85
6.60
6.30
0.10
K=32
8.97
3.07
114.22
0.08
10.21
5.11
7.03
0.12
Table 2: Ablation on Time Encoding’s Token Size ( K ).
Setting
KITTI 00–10 (10Hz)
nuScenes (12Hz)
terr
rerr
ATE
serr
terr
rerr
ATE
serr
{12/6/4}
10.24
4.24
152.42
0.09
9.38
4.02
6.15
0.11
{12} ∗
12.89
4.40
213.76
0.09
10.23
4.91
6.35
0.12
{12/6}
8.64
2.91
117.58
0.07
8.08
3.40
5.27
0.10
{12/6/4}
8.97
3.43
93.23
0.05
9.04
3.86
5.91
0.10
{12/6/4/3}
9.43
3.57
97.21
0.06
9.06
4.21
6.02
0.11
Table 3: Ablation on Training Observation Frequency. Shaded row : OpenVO trained without proposed Time Condition Layers; Other rows : OpenVO trained with proposed Time Condition Layers. ∗ indicates that OpenVO uses 12 Hz as the fixed inference temporal condition, regardless of the dataset.
Setting
OpenVO
ZeroVO ‡ [ 23 ]
terr
rerr
ATE
terr
rerr
ATE
KITTI (2.5Hz)
28.83
8.22
368.47
52.74
17.46
553.52
KITTI (5Hz)
12.06
3.77
132.72
28.62
8.78
387.10
nuScenes (3Hz)
12.73
3.71
7.60
25.75
8.30
14.64
nuScenes (4Hz)
8.14
3.24
5.00
15.49
6.08
9.33
nuScenes (6Hz)
9.18
4.05
6.07
31.24
10.30
21.55
Table 4: Ablation on Varying Inference Observation Rates. ‡ indicates ZeroVO uses the same Metric Depth [ 17 ] and Intrinsic [ 68 ] priors from foundation models as OpenVO. Green cells denote OpenVO (left) improved performance of over existing method (right) . Our OpenVO is significantly more robust to observation rate variations, consistently reducing VO errors by 46% to 92% across different metrics and benchmark settings.
Method
# params (M)
Runtime (s)
WildCamera [ 68 ]
270.447
41.906
Metric3Dv2 [ 17 ]
411.941
180.752
Table 5: Expected latency of foundation priors.
Method
# params (M)
Runtime (s)
Flow Estimator
20.655
3.089
Time Condition Layers
0.003673
0.0101
Diff. 2D-Guided 3D Flow
-
0.0204
Self-Attn
174.695
0.0339
Context Encoder
56.702
0.0129
Egomotion Decoder
29.988
0.4368
Table 6: Expected latency of our OpenVO. Red rows indicate Time-Aware Flow Encoder block. Green row indicates Geometry-Aware Context Encoder block. Purple row indicates Egomotion decoder block.
Setting
KITTI 00–10 (10Hz)
nuScenes (12Hz)
terr
rerr
ATE
serr
terr
rerr
ATE
serr
Ours–Oracle
7.22
2.75
90.85
0.05
6.41
3.44
4.24
0.09
Table 7: Oracle testing. We equip OpenVO with ground-truth intrinsic parameters to isolate errors from calibration, revealing the upper-bound performance under perfect camera parameters estimation.
Time-Conditioning Layers
P.E.
terr
rerr
ATE
Cross-Attn (PE( Δt ) in Eq.1)
✓
12.74
4.22
155.30
Cross-Attn (single scalar Δt )
✗
14.74
4.78
180.16
Table 8: Time-Conditioning Layers design. Using Cross-Attention layers as alternatives to the design of Time-Conditioning Layers and the use of our proposed positional encoding (P.E.).
Ablation
P.E.
Flow
Geometry
terr
rerr
ATE
OpenVO
✓
✓
✗
17.12
4.42
158.67
OpenVO
✓
✗
✓
-
-
-
OpenVO
✗
✓
✓
9.59
3.54
117.50
Table 9: Time-Conditioning Layers design. Using Cross-Attention layers as alternatives to the design of Time-Conditioning Layers and the use of our proposed positional encoding (P.E.).
Methods
terr
rerr
ATE
PnP +Metric3Dv2+MaskFlow
54.38
2.81
302.74
DroidCalib [ 15 ] : no Sim(3) ( fair )
69.16
0.625
382.65
DroidCalib [ 15 ] : w/ Sim(3) ( unfair )
11.01
0.625
73.18
ViPE [ 18 ] : no Sim(3) ( fair )
13.29
0.721
137.15
ViPE [ 18 ] : w/ Sim(3) ( unfair )
10.15
0.721
72.73
Table 10: Geometry-based and optimization-based methods.
Figure 4: Modified VectorMapNet [ 28 ] . A front-view input image is first processed by an image encoder to extract semantic and geometric features. These features are then lifted into a bird’s-eye-view (BEV) representation using inverse perspective mapping, which leverages the camera’s intrinsic and extrinsic parameters from OpenVO to geometrically project image features onto the ground plane. The resulting BEV feature map is fed into the Vector Map Decoder, which predicts structured map elements in an intermediate representation consisting of key components such as polyline classes, control points, and geometric attributes. Finally, a polyline generator converts these decoded components into continuous vectorized map elements, such as lane boundaries, road dividers, and crosswalks – yielding a high-resolution, topologically meaningful HD map suitable for downstream driving tasks.
Figure 5: Qualitative results of Global HDMap reconstruction results produced by OpenVO + modified monocular VectorMapNet [ 28 ] . Local mapping outputs are gradually fused through OpenVO’s ego-to-world pose estimates, producing a coherent global HD-map reconstruction of the full scenario. We would like to refer to our supplementary videos for further details of the OpenVO-enabled monocular-based global map reconstruction.
Figure 6: Qualitative HD Map reconstruction results produced by modified monocular VectorMapNet [ 28 ] . The leftmost column displays the input RGB frames; the second column shows the ground-truth HDMaps; the third column presents the results from our modified VectorMapNet using only a single front-camera image; and the last column shows the outputs under the six-camera + LiDAR configuration. We highlight the differences in each example using dashed pink regions .
Figure 7: Qualitative of Stereo benchmark on Argoverse 2 [ 56 ] . Each row shows one example, including the input stereo image and the reference metric depth. The stereo images in Argoverse 2 often provides low-quality or weakly constrained metric depth due to limited disparity in long-range regions and visually challenging street scenes. This degradation leads to information loss and introduces uncertainty into downstream VO estimation.
Figure 8: Qualitative results on KITTI [ 13 ] benchmark. Each row presents one example. The KITTI camera provides a wider field of view than most datasets, allowing it to capture a richer set of dynamic objects while still preserving its long-range odometry characteristics.
Figure 9: Qualitative results on real-world captured videos. We present two examples, each accompanied by the corresponding RGB frames and reference metric-depth images. Real-world videos commonly exhibit numerous environmental artifacts—such as noise, clutter, and dynamic elements—which pose significant challenges for generalizability and real-world performance assessment.
Metrics
Full (6) cams + Lidar
Front cam only
mAP @ 0.5 ( ↑ )
0.187
0.102
mAP @ 1.0 ( ↑ )
0.479
0.318
mAP @ 1.5 ( ↑ )
0.646
0.472
Overall mAP ( ↑ )
0.4375
0.297
mAP ped @ 1.5 ( ↑ )
0.644
0.469
mAP div @ 1.5 ( ↑ )
0.642
0.509
Table 11: Comparison between the full-sensor configuration and our monocular front-camera–only configuration for HDMap reconstruction. Green cells highlight our configuration. ped denotes pedestrian crossings, div denotes dividers, and cont denotes contours. Performance is reported in terms of mean Average Precision (mAP).
Visual odometry is essential for autonomous navigation in GPS-denied environments, yet RGB-based methods remain vulnerable to motion blur, challenging illumination, and dropped frames. Event cameras complement conventional cameras with high temporal resolution and dynamic range, but their asynchronous measurements complicate reliable correspondence estimation. We present DAPEVO, a learned visual odometry system that estimates image and event correspondences independently at shared patch locations and fuses their correlation evidence before motion refinement. Each tracked patch maintains image and event descriptors, and a learned scalar gate combines modality-specific correlation embeddings for each patch--frame edge before a shared recurrent refinement and bundle-adjustment update. DAPEVO also supports event-only observations, enabling continued tracking when RGB frames are sparse or unavailable, while modality-aware keyframe culling preserves scarce frame constraints. On UZH-FPV, when retaining only one in six RGB frames, DAPEVO's mean absolute trajectory error (ATE) increases by only 36%, from 1.00 to 1.36m, whereas the ATE of DPVO and RAMP-VO rises by factors of 3.7× and 3.1×, respectively. On TartanEvent, DAPEVO similarly remains below 1m ATE at 3Hz RGB input, while DPVO and RAMP-VO exceed 9m. Under degraded RGB input on TartanEvent, DAPEVO achieves an ATE of 0.60m, compared with more than 4m for both DPVO and RAMP-VO, while also outperforming event-only DEVO at 0.87m.
Stable and reliable 4D spatial understanding is fundamental for autonomous driving systems. While feedforward reconstruction networks can estimate camera motion and 3D structure in one pass, pose estimation over long videos remains challenged by computational cost, long-context ambiguity, and temporal instability. To address these challenges, we propose Feedforward Visual Odometry (FFVO), a pose-specialized adaptation of joint reconstruction architectures for efficient and temporally stable camera-pose estimation. FFVO uses (i) a compact camera-token representation for computationally efficient temporal aggregation, (ii) a hierarchical local-to-global temporal decoder that mitigates geometric ambiguity by separating short-range motion aggregation from sequence-level integration, and (iii) intermediate trajectory supervision that promotes temporal stability. Extensive evaluation on the Waymo Open Dataset (WOD), KITTI, and a large-scale proprietary benchmark demonstrates that our method performs favorably against existing feedforward approaches, and greatly reduces jitter and drift. These results support FFVO as an effective feedforward camera-pose decoder in long-horizon visual odometry settings.
Meng-Li Shih, Shih-Yang Su, Yuliang Zou +7
Waymo LLC · University of Washington · Google DeepMind
The most accurate monocular visual odometry systems require known camera intrinsics, refine their estimates with test-time optimization, and recover trajectories only up to an unknown factor. Systems built on large 3D models need no intrinsics, but they remain considerably less accurate and slower for odometry. Direct pose regression avoids all these requirements, yet it has not matched either approach's accuracy. We revisit this formulation with a transformer that predicts relative camera poses together with separate rotation and translation confidences over overlapping image windows, supervised by camera poses alone. A confidence-weighted module then aggregates the overlapping predictions into a single trajectory. The resulting method, CalfVO, needs no intrinsics, no bundle adjustment, and no loop closure, and it recovers scale from learned priors, accurately enough that it is evaluated without any alignment to the ground truth. Across five benchmarks, it is the most accurate calibration-free method on every metric we report, and it runs at 53 FPS, faster than every baseline.