acceptodds
Under review as a conference paper at ICLR 2027

Drive3R: Block-wise Streaming Geometric Foundation Models for Autonomous Driving

Abstract

High-fidelity closed-loop simulation for autonomous driving increasingly relies on dense multi-view 3D reconstruction. While recent feed-forward geometric foundation models have revolutionized per-scene optimization, applying them to continuous, long-horizon driving sequences is computationally intractable due to GPU memory limits and the quadratic complexity of global attention. Existing scaling strategies are fundamentally limited, as causal streaming suffers from compounding scale drift and trajectory deviation, while rigid "chunk-then-align" heuristics fragment global context and produce trajectory discontinuities. To address these limitations, we introduce Drive3R, a block-wise streaming geometric foundation model. Its block-wise causal attention resolves the scalability bottleneck of fully bidirectional attention while avoiding the loss of local future context in purely causal streaming. Its confidence-aware re-anchoring further mitigates long-range drift by detecting unreliable streaming states, re-initializing the KV cache, and stitching refreshed segments through alignment. Extensive experiments demonstrate that Drive3R establishes a new state-of-the-art on nuScenes, KITTI, and Waymo. Specifically, on nuScenes, it improves over the recent DVGT-2 baseline from 0.1806 to 0.1081 in AbsRel, from 0.6750 to 0.5344 in point-cloud Accuracy, and from 1.5812 to 0.9092 in ATE. Code will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Autonomous Driving, Geometric Foundation Model, Long-Horizon 3D Reconstruction, Block-wise Streaming, Confidence-Aware Re-Anchoring

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_8QuAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Drive3R, a block-wise streaming geometric foundation model for autonomous driving. It combines intra-block bidirectional attention, cross-block KV caching, and confidence-aware re-anchoring to improve long-horizon 3D reconstruction. The experimental results demonstrate strong performance across nuScenes, Waymo, and KITTI.

Critical strengths
  1. The paper addresses an important and practical challenge in scalable geometric reconstruction for autonomous driving.
  2. The proposed framework is well-motivated, with a clear architecture and effective integration of temporal context and drift mitigation.
Critical weaknesses
  1. Block-wise causal attention, KV caching, and confidence-triggered re-initialization are individually established techniques. The main contribution appears to be their integration. More controlled comparisons are needed to establish the novelty and benefits beyond existing streaming and chunk-based approaches.

  2. The proposed SE(3) re-anchoring cannot directly correct accumulated scale inconsistencies, while KITTI trajectory evaluation relies on Sim(3) alignment. This raises questions about whether the method genuinely addresses long-term metric-scale drift.

  3. The KV cache continues growing between resets, making computational scalability dependent on the re-anchoring frequency. Moreover, Drive3R has higher inference latency than DVGT-2 (21.10s vs. 12.64s), and the efficiency analysis is limited to relatively short sequences.

Questions for authors
  1. Can the authors provide more controlled comparisons to distinguish the contribution of block-wise attention from confidence-aware re-anchoring, using matched architectures and reset budgets?

  2. How does SE(3) re-anchoring mitigate accumulated metric-scale drift? Could the authors report unaligned long-horizon trajectory errors and scale-drift measurements?

  3. How do inference latency, peak memory, and reset frequency scale with substantially longer sequences?

3 comments

Response to Reviewer 8QuA

We thank the reviewer for the careful evaluation and for recognizing the practical importance of scalable geometric reconstruction for autonomous driving.

Response 1: Problem-driven design and controlled comparison

Reviewer's Comment (Critical Weakness 1 / Question 1): Block-wise attention, KV caching, and confidence-triggered reinitialization are individually established techniques. More controlled comparisons are needed to distinguish their contributions using matched architectures and reset budgets.

Response:

Long-horizon surround-view reconstruction in autonomous driving imposes coupled requirements that generic sequence models do not directly address: local geometry must remain consistent across a calibrated multi-camera rig, useful context must persist across blocks, accumulated trajectory and scale drift must be removed before it dominates future predictions, and streaming memory must remain bounded. Generic frame-wise and chunk-based pipelines address only parts of this problem. Drive3R addresses these coupled requirements through a learnable streaming framework, rather than relying on hand-designed long-range heuristics.

The design follows directly from these requirements. Block-wise attention constructs a local multi-camera geometric state; KV memory preserves useful history across blocks; a confidence head trained from trajectory drift determines when that state should be refreshed; and SE(3) alignment reconnects the refreshed segment to the global trajectory. These mechanisms form a coherent state-update pipeline rather than a stack of unrelated generic operations.

Two controlled analyses in the paper isolate local bidirectional context and confidence-aware re-anchoring. We further add a reset-budget-matched comparison to evaluate adaptive reset timing.

ComparisonWhat is isolatedEvidence
Block size K=1K=1 to K=12K=12Effect of intra-block bidirectional contextFigure 6
With vs. without confidence-aware re-anchoringEffect of learned resettingTable 4

Figure 6 isolates the effect of block-wise context by changing only the block size under the same backbone. Table 4 keeps the block-wise setting and removes confidence-aware re-anchoring; this causes ATE to increase from 4.1422 to 18.1769.

Table R1: Controlled reset-policy experiment on 20 nuScenes scenes

Reset policyMean resets per sceneMean cache length (frames)P95 cache length (frames)Forward time (s)Peak GPU memory (GB)ATE Sim(3) (m) ↓AbsRel ↓
No reset0.0016.503315.66625.0013.4560.2788
Fixed interval, closest reset count (9-frame interval)4.002.79611.76114.300.6170.1405
Fixed interval, closest forward time (12-frame interval)3.004.15912.03215.491.1760.1631
Confidence-based, tau = 0.93.803.791212.06316.130.4180.1351

The two fixed-interval baselines provide closely matched reset counts or forward time. The confidence-based policy uses 3.80 resets per scene versus 4.00 for the closest reset-count baseline, while reducing ATE from 0.617 to 0.418. At nearly identical forward time (12.063 versus 12.032 s), it reduces ATE from 1.176 to 0.418.

Together, these analyses separate the effect of local bidirectional context from that of learned state resets and show that the latter is not simply due to more frequent resetting.

Response 2: Metric-scale consistency without Sim(3) correction

Reviewer's Comment (Critical Weakness 2 / Question 2): SE(3) re-anchoring cannot directly correct accumulated scale inconsistencies, while KITTI evaluation uses Sim(3) alignment. How can metric-scale drift be evaluated without Sim(3) correction?

Response:

The reviewer identifies an important distinction. In Drive3R, each reinitialized segment predicts its own metric scale through the metric-scale branch; scale is neither recovered from nor corrected by the SE(3) stitching transform. The predicted segment is then placed into the global frame using the six-DoF transform estimated on overlapping poses.

Unlike relative-scale models such as VGGT, which are trained without explicit metric-scale supervision, Drive3R directly supervises metric scale during training. We used Sim(3) alignment in the main KITTI table because several compared methods output trajectories only up to relative scale. To evaluate Drive3R independently of this scale-invariant protocol, we report the following metrics:

  • Unaligned ATE: compares predicted and ground-truth trajectories without any global alignment.
  • SE(3)-aligned ATE: allows global rotation and translation but no scale correction.
  • Global scale drift: measures the residual metric-scale discrepancy.
  • Boundary RPE: measures local relative-motion error around reset events.

Table R2: Metric-scale consistency and boundary continuity on complete KITTI

MethodUnaligned ATE (m) ↓SE(3)-aligned ATE (m) ↓Median global scale error (%) ↓Common-window translation RPE (m) ↓Common-window rotation RPE (deg) ↓
DA3-Streaming278.015150.73484.0216.5743.697
DVGT-2190.206104.91118.503.8274.195
Drive3R93.12149.4835.132.7001.033

Without any global alignment, Drive3R reduces ATE by 51.0% relative to DVGT-2 and by 66.5% relative to DA3-Streaming. Under SE(3)-only alignment, the corresponding reductions are 52.8% and 67.2%. Its median global scale error is 5.13%, compared with 18.50% for DVGT-2 and 84.02% for DA3-Streaming.

All three methods use the same fixed reference reset events and temporal windows. Each window contains three frames before and three frames after a Drive3R reset event, yielding 717 valid windows. The RPE therefore measures local relative-motion error over identical intervals; it does not require DVGT-2 to reset at these timestamps or the DA3-Streaming windows to coincide with its chunk boundaries.

On these common windows, Drive3R achieves translation and rotation RPEs of 2.700 m and 1.033 degrees, lower than DVGT-2 (3.827 m and 4.195 degrees) and DA3-Streaming (16.574 m and 3.697 degrees). These results demonstrate lower metric-scale drift and better local consistency across resets without relying on Sim(3) scale correction.

Response 3: Efficiency and scaling on complete sequences

Reviewer's Comment (Critical Weakness 3 / Question 3): The KV cache grows between resets, and the efficiency analysis is limited to short sequences. How do latency, peak memory, and reset frequency scale with substantially longer sequences?

Response:

Table R3 compares runtime and peak GPU memory for DA3-Streaming, DVGT-2, and Drive3R on all 11 complete KITTI sequences, which span 55 to 933 frames and 0.4 to 5.1 km. The longest sequence is 25.9 times longer than the 36-frame efficiency clip. All three methods complete the evaluation without OOM. The internal closed-course sequences in Appendix C.2 also complete without OOM.

Table R3: Efficiency on complete KITTI sequences

MethodOOM sequencesMean inference + stitching time (s)P95 sequence time (s)Amortized time (s / stereo timestamp)Mean per-sequence peak GPU memory (GiB)
DA3-Streaming0145.502329.5060.344320.09
DVGT-2031.25168.1500.073919.81
Drive3R038.41184.1610.090913.60

DVGT-2 uses factorized spatial-temporal attention, which separates cross-view spatial attention from temporal attention and reduces the number of tokens involved in each attention operation. DA3-Streaming follows the chunk-then-align paradigm. Drive3R retains joint cross-view and temporal interactions; applying similar factorization reduced reconstruction quality in our experiments.

On complete KITTI logs, Drive3R takes 38.411 s on average, 22.9% longer than DVGT-2 (31.251 s) but 73.6% shorter than DA3-Streaming (145.502 s). Drive3R has the lowest mean peak GPU memory: 13.60 GiB, compared with 19.81 GiB for DVGT-2 and 20.09 GiB for DA3-Streaming, corresponding to reductions of 31.3% and 32.3%.

Across complete KITTI logs, Drive3R performs 65.45 resets per sequence on average. The P95 and global maximum cached-history lengths are 6 and 21 frames, respectively. The revised version will include an optional defensive cache cap B_max, which forces a reset when the cache reaches the bound even if confidence remains above the threshold. This safety guard is not enabled in the experiments reported here.

Taken together, Tables R2 and R3 show a favorable accuracy-efficiency trade-off: Drive3R completes all evaluated long sequences without OOM and uses substantially less average peak memory than both baselines, while achieving lower SE(3)-aligned ATE, global scale error, and boundary RPE. Its runtime is moderately higher than DVGT-2 but substantially lower than DA3-Streaming.

Reviewer_h71dOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents Drive3R, a block-wise streaming geometric foundation model for long-horizon 3D reconstruction in autonomous driving. The proposed framework combines block-wise causal attention with confidence-aware re-anchoring to address the limitations of fully bidirectional reconstruction and conventional causal streaming methods. Specifically, the block-wise attention mechanism enables bidirectional interactions within local temporal blocks while maintaining historical information through cross-block KV caching. Furthermore, the confidence-aware re-anchoring strategy adaptively resets unreliable streaming states and aligns reconstructed segments to mitigate accumulated trajectory drift. Extensive experiments on nuScenes, Waymo, and KITTI demonstrate competitive performance in depth estimation, point-cloud reconstruction, and ego-motion estimation.

Critical strengths
  1. A practical and well-motivated streaming reconstruction framework. The proposed framework effectively combines intra-block bidirectional attention with cross-block KV caching, balancing local geometric consistency and long-horizon streaming efficiency. This design addresses the limitations of both per-frame streaming and independent chunk-based reconstruction.
  2. Strong experimental performance across multiple benchmarks. The proposed method achieves strong performance in depth estimation, point-cloud reconstruction, and ego-motion estimation across nuScenes, Waymo, and KITTI. The experimental results demonstrate its effectiveness and potential for long-horizon autonomous driving reconstruction.
Critical weaknesses

The proposed framework combines block-wise attention, KV-cache propagation, confidence-based resets, and rigid alignment. While their integration is effective, several components are closely related to established techniques for long-sequence processing and geometric alignment. The paper could better clarify the specific methodological novelty beyond combining these existing ideas.

The re-anchoring strategy uses SE(3) alignment, which corrects rigid pose discrepancies but does not explicitly correct scale differences between reconstructed segments. Moreover, the long-sequence trajectory evaluation on KITTI reports Sim(3)-aligned ATE. While the reported improvements are encouraging, additional analysis of metric-scale consistency and unaligned trajectory drift would provide stronger evidence for the claimed long-horizon geometric stability.

Questions for authors

Could the authors explain how scale consistency is maintained across reconstructed segments when the re-anchoring mechanism only performs SE(3) alignment? Additionally, could the authors provide an evaluation of metric-scale drift without Sim(3) alignment to further validate long-horizon geometric consistency?

4 comments

Response to Reviewer h71d

We thank the reviewer for the careful evaluation and for recognizing the practical value of Drive3R and its strong performance across nuScenes, Waymo, and KITTI.

Response 1: Problem-specific design for long-horizon driving reconstruction

Reviewer's Comment (Critical Weakness 1): Block-wise attention, KV caching, and confidence-based resets are individually established techniques. The paper should more clearly identify the methodological novelty beyond combining these components.

Response:

Long-horizon surround-view reconstruction in autonomous driving imposes coupled requirements that generic sequence models do not directly address: synchronized multi-camera observations must remain geometrically consistent, pose and metric-scale errors must be controlled over long trajectories, and historical features must remain within a bounded streaming budget. Drive3R addresses this problem through a learnable streaming framework, rather than relying on hand-designed long-range heuristics.

The method-specific design follows from these requirements:

  1. Multi-camera local state construction. Each attention block spans multiple synchronized frames and cameras, allowing local bidirectional context across the driving rig. This differs from frame-wise causal models that treat the stream as a generic sequence of independent tokens.
  2. Short-term continuity across consecutive blocks. Historical geometric features are carried through KV memory, so adjacent blocks remain connected instead of being reconstructed as independent chunks.
  3. Long-term state refresh and geometric reconnection. A confidence head trained from block-level trajectory drift predicts when the current geometric state becomes unreliable. The model then discards the stale cache, starts a new local segment, and uses SE(3) alignment only to reconnect that refreshed segment to the global trajectory. Unlike fixed chunk boundaries, the reset location is determined adaptively.

The contribution is the problem-specific coupling of cross-block feature propagation with confidence-driven state refresh and geometric reconnection. This design directly addresses the reliability-continuity trade-off in long-horizon driving reconstruction: useful history remains available while the state is reliable, and invalid history is replaced before it dominates future predictions.

Response 2: Metric-scale consistency without Sim(3) scale correction

Reviewer's Comment (Critical Weakness 2 / Question): SE(3) alignment does not correct scale differences, while KITTI ATE uses Sim(3) alignment. How is metric-scale consistency maintained across segments, and can the authors report unaligned long-horizon trajectory errors and scale drift?

Response:

The reviewer identifies the key distinction correctly. In Drive3R, each reinitialized segment predicts its own metric scale through the metric-scale branch; scale is neither recovered from nor corrected by the SE(3) stitching transform. The predicted segment is then placed into the global frame using the six-DoF transform estimated on overlapping poses. Unlike relative-scale models such as VGGT, which are trained without explicit metric-scale supervision, Drive3R directly supervises metric scale during training. Segment scale therefore originates from the model prediction and training objective rather than post-hoc alignment.

Our previous KITTI evaluation uniformly applies Sim(3) alignment because several compared models, including VGGT-style methods, output trajectories only up to relative scale. This provides a scale-invariant comparison across methods with different scale conventions, but it cannot by itself establish metric-scale stability. We therefore evaluate the following metrics without Sim(3) scale correction:

  • Unaligned ATE: compares the predicted and ground-truth trajectories without any global alignment.
  • SE(3)-aligned ATE: allows global rotation and translation but no scale correction.
  • Global scale drift: measures the residual metric-scale discrepancy.
  • Boundary RPE: measures local relative-motion error around reset events.

Table R1: Metric-scale consistency and boundary continuity on complete KITTI

MethodUnaligned ATE (m) ↓SE(3)-aligned ATE (m) ↓Median global scale error (%) ↓Common-window translation RPE (m) ↓Common-window rotation RPE (deg) ↓
DA3-Streaming278.015150.73484.0216.5743.697
DVGT-2190.206104.91118.503.8274.195
Drive3R93.12149.4835.132.7001.033

Without any global alignment, Drive3R reduces ATE by 51.0% relative to DVGT-2 and by 66.5% relative to DA3-Streaming. Under SE(3)-only alignment, the corresponding reductions are 52.8% and 67.2%. Its median global scale error is 5.13%, compared with 18.50% for DVGT-2 and 84.02% for DA3-Streaming.

All three methods use the same fixed reference reset events and temporal windows. Each window contains three frames before and three frames after a Drive3R reset event, yielding 717 valid windows. The RPE therefore measures local relative-motion error over identical intervals; it does not require DVGT-2 to reset at these timestamps or the DA3-Streaming windows to coincide with its chunk boundaries.

Drive3R achieves translation and rotation RPEs of 2.700 m and 1.033 degrees, lower than DVGT-2 (3.827 m and 4.195 degrees) and DA3-Streaming (16.574 m and 3.697 degrees). These results demonstrate lower metric-scale drift and better local consistency across resets without relying on Sim(3) scale correction.

Thank you for the detailed response and additional experiments.

The clarification of the confidence-aware re-anchoring mechanism helps explain the motivation behind the proposed streaming framework. The reported improvements in unaligned ATE, SE(3)-aligned ATE, scale error, and boundary RPE provide stronger evidence for the geometric stability of Drive3R.

I appreciate the authors' efforts in addressing my concerns. Given these additional results, I am updating my rating from 3 to 4.

Dear Reviewer h71d,

Thank you very much for your positive feedback and for raising the score! We truly appreciate your time and constructive comments throughout the review process.

Best regards, The Authors

Reviewer_1xW2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Drive3R addresses streaming 3D reconstruction and ego-motion estimation from surround-view driving videos. It introduces block-wise causal attention: frames within a block interact bidirectionally, while information from earlier blocks is retained through a KV cache. A confidence branch estimates streaming reliability and triggers cache resets when confidence falls below a threshold. Newly initialized segments are aligned with the existing reconstruction using overlapping poses and an SE(3) transformation. Evaluation includes depth and point-cloud reconstruction on nuScenes and Waymo, long-sequence trajectory estimation on KITTI, and supporting analyses of robustness and efficiency.

Critical strengths
  1. The proposed architecture is well suited to multi-camera streaming reconstruction. Block-wise attention retains local bidirectional context while allowing historical features to propagate across blocks. The strong depth and point-cloud results on nuScenes and Waymo, together with improved average trajectory accuracy on KITTI, demonstrate the practical value of this design.

  2. The supplementary experiments provide useful evidence for the design choices. The paper examines confidence–drift correlations, fixed-interval resetting, training from scratch, calibration perturbations, and generalization to a different camera rig. These analyses strengthen the empirical evaluation. In particular, the confidence–drift correlation supports using the learned confidence as a reliability signal.

Critical weaknesses
  1. The contribution of confidence-based reset timing remains unclear. Table 9(a) shows an improvement over fixed-interval resetting, but the reset counts, average cache lengths, and computational costs of the two policies are not reported. The selection of the fixed interval also needs clarification. Moreover, increasing the confidence threshold consistently improves performance in Table 9(b), potentially through more frequent resets. A comparison with a tuned fixed-interval policy under matched reset counts or computational budgets is therefore needed to establish the benefit of adaptive timing.

  2. The long-sequence evaluation does not fully support the claims about scalability and metric-scale stability. The cache in Eq. (4) grows until a reset is triggered, and the paper does not specify a maximum cache length or fallback mechanism. Runtime and memory measurements on 36-frame clips provide limited evidence about scaling to complete driving logs. In addition, the main long-sequence trajectory results use Sim(3) alignment, which removes global scale error. Since SE(3) stitching cannot correct scale discrepancies between segments, additional measurements are needed to establish scale stability and local motion continuity across reset boundaries.

Questions for authors
  1. How was the fixed reset interval in Table 9(a) selected? Please report reset counts and average cache lengths, and clarify whether confidence-based resetting remains better when reset counts or computational budgets are matched.

  2. Is cache growth bounded if confidence remains above the threshold? What are the peak memory usage, runtime, and reset statistics on complete KITTI sequences?

  3. Can you report SE(3)-aligned trajectory errors, scale errors, and relative pose errors around reset boundaries? These measurements would clarify whether single-frame overlap is sufficient to maintain metric scale and motion continuity.

2 comments

Response to Reviewer 1xW2

We thank the reviewer for the careful evaluation and for recognizing the value of block-wise streaming, confidence-aware re-anchoring, and the additional analyses.

Response 1: Adaptive reset timing under matched reset and latency budgets

Reviewer's Comment (Weakness 1 / Question 1): The improvement over fixed-interval resetting may be confounded by differences in reset count, cache length, and computation cost. A tuned fixed-interval baseline with matched reset counts or computational budgets is needed to establish the benefit of adaptive timing.

Response:

To isolate reset placement from reset frequency and computational cost, we conduct a controlled comparison on 20 nuScenes scenes. Each scene contains 36 frames and six cameras, and all policies use the same input, model weights, and inference configuration.

The fixed-interval baselines are selected to closely match either the average reset count or the forward time of the confidence-based policy.

Table R1: Reset-policy comparison on 20 nuScenes scenes

Reset policyMean resets per sceneMean cache length (frames)P95 cache length (frames)Forward time (s)Peak GPU memory (GB)ATE Sim(3) (m) ↓AbsRel ↓
No reset0.0016.503315.66625.0013.4560.2788
Fixed interval, closest reset count (9-frame interval)4.002.79611.76114.300.6170.1405
Fixed interval, closest forward time (12-frame interval)3.004.15912.03215.491.1760.1631
Confidence-based, tau = 0.93.803.791212.06316.130.4180.1351

Compared with the reset-count-matched fixed policy, the confidence-based policy uses fewer resets on average (3.80 versus 4.00), while reducing ATE from 0.617 to 0.418 (32.2%) and AbsRel from 0.1405 to 0.1351.

Compared with the latency-matched fixed policy, the forward times are nearly identical (12.063 versus 12.032 s), while the confidence-based policy reduces ATE from 1.176 to 0.418 (64.5%) and AbsRel from 0.1631 to 0.1351 (17.2%).

These results show that the improvement is not merely due to more frequent cache clearing or a larger computational budget. The learned reset locations remain beneficial under both reset-count and latency controls.

Response 2: Full-sequence scalability and computational cost

Reviewer's Comment (Weakness 2 / Question 2): The KV cache grows between resets, and the paper reports efficiency only on relatively short clips. Runtime, memory, and reset statistics are needed on complete KITTI sequences.

Response:

Table R2 compares runtime and peak GPU memory for DA3-Streaming, DVGT-2, and Drive3R on all 11 complete KITTI sequences, which span 55 to 933 frames and 0.4 to 5.1 km. All three methods complete the evaluation without OOM. The internal closed-course sequences in Appendix C.2 also complete without OOM.

Table R2: Efficiency on complete KITTI sequences

MethodOOM sequencesMean inference + stitching time (s)P95 sequence time (s)Amortized time (s / stereo timestamp)Mean per-sequence peak GPU memory (GiB)
DA3-Streaming0145.502329.5060.344320.09
DVGT-2031.25168.1500.073919.81
Drive3R038.41184.1610.090913.60

DVGT-2 uses factorized spatial-temporal attention, which separates cross-view spatial attention from temporal attention. This reduces the number of tokens involved in each attention operation and gives DVGT-2 the lowest latency. DA3-Streaming follows the chunk-then-align paradigm. Drive3R instead retains joint cross-view and temporal interactions. In our experiments, applying similar factorization to Drive3R reduced reconstruction quality, so we preserve the joint attention design.

On complete KITTI logs, Drive3R takes 38.411 s on average, 22.9% longer than DVGT-2 (31.251 s) but 73.6% shorter than DA3-Streaming (145.502 s).

Drive3R has the lowest mean peak GPU memory among the evaluated streaming methods: 13.60 GiB, compared with 19.81 GiB for DVGT-2 and 20.09 GiB for DA3-Streaming, corresponding to reductions of 31.3% and 32.3%, respectively.

Across complete KITTI logs, Drive3R performs 65.45 resets per sequence on average. The P95 and global maximum cached-history lengths are 6 and 21 frames, respectively. The revised version will include an optional defensive cache cap B_max, which forces a reset when the cache reaches the bound even if confidence remains above the threshold. This safety guard is not enabled in the experiments reported here.

Response 3 and Table R3 further compare trajectory and metric-scale performance on the same sequences, where Drive3R achieves lower SE(3)-aligned ATE, global scale error, and boundary RPE than both baselines.

Response 3: Metric-scale consistency and boundary continuity

Reviewer's Comment (Weakness 2 / Question 3): SE(3) stitching cannot correct scale differences, and KITTI ATE uses Sim(3) alignment. Please report SE(3)-aligned errors, scale errors, and relative pose errors around reset boundaries.

Response:

To compare with methods that do not produce native metric-scale trajectories, our previous KITTI evaluation uniformly applies Sim(3) alignment. Sim(3) alignment removes global scale and therefore cannot by itself establish metric-scale stability. To address this, we compute three metric-space diagnostics from the existing KITTI trajectories without applying Sim(3) scaling: SE(3)-aligned ATE, global scale drift, and relative pose errors (RPE) at reset events.

Table R3: Metric-scale consistency and boundary continuity on complete KITTI

MethodSE(3)-aligned ATE (m) ↓Median global scale error (%) ↓Common-window translation RPE (m) ↓Common-window rotation RPE (deg) ↓
DA3-Streaming150.73484.0216.5743.697
DVGT-2104.91118.503.8274.195
Drive3R49.4835.132.7001.033

Without Sim(3) scale correction, Drive3R reduces SE(3)-aligned ATE by 52.8% relative to DVGT-2 and by 67.2% relative to DA3-Streaming. Its median global scale error is 5.13%, compared with 18.50% for DVGT-2 and 84.02% for DA3-Streaming.

All three methods use the same fixed reference reset events and the same temporal windows. Each window contains three frames before and three frames after a Drive3R reset event; 717 valid windows are evaluated.

Consequently, the reported RPE measures local relative-motion error on these common windows for all three methods, rather than each method's native sliding-window or chunk-boundary error. DVGT-2 does not require a reset to exist at these timestamps, and the DA3-Streaming windows do not need to coincide with its chunk boundaries.

On these common windows, Drive3R achieves translation and rotation RPEs of 2.700 m and 1.033 degrees, lower than DVGT-2 (3.827 m and 4.195 degrees) and DA3-Streaming (16.574 m and 3.697 degrees). These results indicate lower metric-scale drift and less local pose discontinuity after reset under the default one-frame overlap setting.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.