acceptodds
Under review as a conference paper at ICLR 2027

DriveWeave: Feed-Forward Dynamic Scene Geometry for Autonomous Driving

Abstract

Reconstructing dynamic driving scenes requires jointly modeling 3D geometry and its temporal evolution. Existing paradigms typically estimate geometry and motion independently, lacking explicit mechanisms to couple their evolution and resulting in spatiotemporal scale drift and structural misalignment. We present DriveWeave, a unified feed-forward foundation paradigm for autonomous driving that predicts metric depth, optical flow, 6-DoF ego-motion, and 3D Gaussian primitives from surround-view videos in a single pass. To resolve the quadratic complexity of dense 4D attention, we introduce 4D Synchronized Register Attention (4D-SyncAttn), which leverages view-frame register synchronization to enable efficient global context aggregation. Furthermore, we formulate a flow-driven spatiotemporal evolution mechanism that enforces complementary differentiable photometric and geometric constraints across frames, encouraging a unified and temporally consistent representation of geometry, motion, and 3DGS rendering manifolds. Extensive evaluations across diverse driving benchmarks demonstrate that DriveWeave achieves state-of-the-art performance in metric depth estimation, 3D point cloud reconstruction, ego-motion estimation, and gaussian scene rendering with superior feed-forward efficiency.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Feed-Forward Visual Geometry, Autonomous Driving

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_A34LOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DriveWeave, a feed-forward model that jointly predicts metric depth, optical flow, ego-motion, and 3D Gaussians from multi-camera videos.

It introduces (1) 4D-SyncAttn to share information across views and time, and uses (2) flow-guided losses to improve depth and rendering consistency across frames.

Experiments on several driving datasets evaluate depth estimation, ego-motion, point cloud reconstruction, and rendering quality, with ablation studies on the attention design and consistency losses.

Critical strengths

The paper combines depth, optical flow, ego-motion, and Gaussian reconstruction in one feed-forward model. Table 6 shows that the consistency losses improve geometry and rendering results, supporting the value of coupling these tasks.

Critical weaknesses

None

Questions for authors

Major concerns

  1. I’m not sure how the depth consistency loss handles moving objects. Suppose the camera is stationary and a car moves toward it. Its depth should change, even if the depth and flow predictions are both correct. Would Eqs. (1)–(2) penalize this? Is there a mask for moving objects or some additional motion compensation?
  2. I found the rendering setup in Table 4 a bit unclear. Which images does the model see, and which ones does it render for evaluation? Are the target images excluded from the inputs? Do the baselines use the same images and camera information?
  3. Since both consistency losses rely on optical flow, I’d like to know how accurate the flow predictions are. Do you have any quantitative results for this?

Minor concerns

  1. In Table 5, adding synchronized registers makes the model faster, from 1.68 s to 1.17 s. Why is that? Does it replace some attention operations?
  2. The pose AUC tells us about rotation and translation direction, but not whether the predicted travel distance is correct. Do you also measure translation-distance or trajectory errors?
2 comments

We sincerely thank the reviewer for the thorough and constructive feedback. Below, we provide detailed clarifications, mathematical justifications, and additional quantitative evaluations addressing moving-object temporal consistency, rendering setups, optical flow accuracy, register-based latency optimization, and trajectory metric evaluations.

Response to Major Concern 1: Handling of Independently Moving Objects in FDC

Reviewer Concern: I’m not sure how the depth consistency loss handles moving objects. Suppose the camera is stationary and a car moves toward it... Would Eqs. (1)–(2) penalize this? Is there a mask for moving objects or some additional motion compensation?

Our Response: We appreciate this sharp observation. To clarify, FDC does not assume a static background scene or rely solely on rigid camera ego-motion warping; rather, it is anchored entirely by the predicted optical flow field Ft→t+1\mathbf{F}_{t \to t+1}.

  • Why Moving Objects Are Not Erroneously Penalized: Given pixel xt\mathbf{x}_{t} in frame tt, optical flow determines its actual corresponding location in the subsequent frame as xt+1=xt+F(xt)\mathbf{x}_{t+1} = \mathbf{x}_{t} + \mathbf{F}(\mathbf{x}_{t}). For a car moving independently toward a stationary camera, the flow vector F\mathbf{F} accurately captures the object's 2D image-plane expansion and displacement.
  • Depth Consistency via Flow Trajectories: When computing the FDC loss, the predicted depth at xt+1\mathbf{x}_{t+1} is evaluated against the warped depth along the true flow trajectory. Because the flow vector incorporates the object's independent motion, the target depth appropriately tracks the object's new 3D location in frame t+1t+1 (accounting for its changed distance), rather than forcing a rigid-background depth constraint.
  • Role in Optimization: Thus, FDC functions as a soft geometric regularization objective that complements per-frame metric depth supervision, avoiding the need for heuristic semantic masks or separate dynamic motion branches. As demonstrated in Fig. 4 and the Temporal Absolute Error (TAE) improvements in Table 6, FDC preserves sharp, temporally coherent depth boundaries even in complex dynamic traffic.

Response to Major Concern 2: Rendering Setup and Baseline Parity in Table 4

Reviewer Concern: Which images does the model see, and which ones does it render for evaluation? Are the target images excluded from the inputs? Do baselines use the same images and camera information?

Our Response: We apologize for the initial ambiguity and clarify the rendering protocols for Table 4 as follows:

  1. Evaluation Modes in Table 4:
    • Per-View Reconstruction: Each camera image independently predicts a local 3D Gaussian representation, which is subsequently rendered and evaluated against its corresponding source RGB observation.
    • Unified-View Reconstruction: All surround-view images are jointly fused into a unified global Gaussian scene representation and rendered across all surround-calibrated viewpoints.
  2. Input/Target Separation & Baseline Parity: In both settings, the target views evaluated are strictly synchronized with the input observations used for reconstruction. All competing baselines are evaluated under the exact same multi-view data splits, calibrated camera intrinsics/extrinsics, and scene sequences to ensure absolute fairness.

Response to Major Concern 3: Optical Flow Accuracy Evaluation

Reviewer Concern: Since both consistency losses rely on optical flow, I’d like to know how accurate the flow predictions are. Do you have any quantitative results for this?

Our Response: To validate the reliability of our flow backbone, we evaluated optical flow accuracy using standard metrics: AEPE (Average End-Point Error, ↓\downarrow) and Fl-all (percentage of outlier pixels, ↓\downarrow). As shown in Table below, DriveWeave consistently outperforms representative dynamic geometry models (such as Flow3R [1]) across both KITTI and SimBEV benchmarks.

MethodKITTI AEPE(px)↓\downarrowKITTI Fl-all(%)↓\downarrowSimBEV AEPE(px)↓\downarrowSimBEV Fl-all(%)↓\downarrow
Flow3R [1]6.6662.3110.3179.11
DriveWeave3.5941.367.6851.52

[1] Zhongxiao Cong, Qitao Zhao, Minsik Jeon, and Shubham Tulsiani. Flow3r: Factored flow prediction for scalable visual geometry learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.438–447, June 2026.

Response to Minor Concern 1: Latency Reduction via Synchronized Registers (Table 5)

Reviewer Concern: In Table 5, adding synchronized registers makes the model faster, from 1.68 s to 1.17 s. Why is that? Does it replace some attention operations?

Our Response: Yes. Synchronized registers introduce a compact, fixed set of global memory tokens that abstract and redistribute scene context, effectively replacing dense, repetitive factorized view and temporal interactions. This substitution substantially alleviates memory bandwidth bottlenecks, reducing the core attention latency from 1.68s to 1.17s without sacrificing representational capacity.

Response to Minor Concern 2: Translation-Distance and Trajectory Errors

Reviewer Concern: The pose AUC tells us about rotation and translation direction, but not whether the predicted travel distance is correct. Do you also measure translation-distance or trajectory errors?

Our Response: Following the same evaluation protocol in Table 2 of the main paper, we additionally report RTE (Relative Translation Error) and RRE (Relative Rotation Error) in the Table below. Furthermore, as illustrated in Fig. 12 (Appendix in the main paper), we provide direct Bird's-Eye-View (BEV) trajectory distance comparisons against ground truth across both short and long horizons, confirming that DriveWeave maintains high metric trajectory scale accuracy without drift accumulation.

MethodNuScenes RTE(cm)↓\downarrowNuScenes RRE(∘^\circ)↓\downarrowDDAD RTE(cm)↓\downarrowDDAD RRE(∘^\circ)↓\downarrowWaymo RTE(cm)↓\downarrowWaymo RRE(∘^\circ)↓\downarrow
CUT3R [1]103.920.7136.310.6131.750.40
DA3 [2]29.260.275.930.066.870.07
VGGT-Ω\Omega [3]43.400.2212.030.069.110.08
DVGT [4]9.390.206.860.118.050.09
DriveWeave6.810.172.900.051.700.05

[1] Wang et al. Continuous 3d perception model with persistent state. CVPR 2025.

[2] Lin et al. Depth anything 3: recovering the visual space from any views. arXiv preprint 2025.

[3] Wang et al. VGGT-Ω. CVPR 2026.

[4] Zuo et al. Dvgt: Driving visual geometry transformer. CVPR 2026.

We sincerely appreciate your valuable feedback and time. Should any further concerns or questions arise during the discussion phase, we are happy to provide additional clarifications. Thank you once more for your constructive input, and we welcome any additional comments you may have.

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

This paper proposes DriveWeave, a feed-forward model that jointly predicts metric depth, optical flow, ego-motion, and 3D Gaussians from multi-frame surround-view images. Its main components are synchronized register attention for efficient spatiotemporal interaction and flow-guided depth and rendering consistency objectives. Training combines synthetic data, public driving datasets, and six million self-collected frames. Experiments evaluate depth, pose, point cloud reconstruction, and Gaussian rendering across multiple driving benchmarks.

Critical strengths
  1. The synchronized register design is interesting. Exchanging compact global tokens across views and timestamps offers a practical way to improve coordination between factorized attention streams without applying dense attention to all image tokens. Relative to the factorized-attention baseline, Table 5 reports improvements in both accuracy and latency.
  2. The shared geometry, motion, and rendering framework is supported by broad experiments. Flow-based cross-task supervision provides a useful connection between the prediction heads, and the consistency-loss ablations show complementary benefits for temporal geometry and rendering quality.
Critical weaknesses
  1. The depth-consistency constraint needs a clearer justification for independently moving objects. Equations (1)-(2) use optical flow to establish correspondences but transform target points using only ego-motion. For an object moving independently along the depth direction, this rigid transformation does not generally recover its earlier depth. The stated validity checks address visibility and correspondence reliability but do not explicitly explain how such motion is excluded or compensated. The loss may therefore impose incorrect consistency targets in precisely the dynamic regions emphasized by the paper.
  2. Training-data comparability with the baselines is unclear. DriveWeave uses six million self-collected frames alongside synthetic and public data and pretrained-model supervision. However, the paper does not state whether the competing methods were retrained or fine-tuned using the same data. The architectural and loss ablations are useful, but if training data differ across methods, the cross-method gains may reflect both methodological and data differences. Clarifying the baseline training protocols, with a controlled comparison where needed, would help assess the contribution.

Some architectural and timing details also require clarification. Table 5 reports a latency reduction from 1.68 s to 1.17 s with synchronized registers, but does not specify the exact layer schedule or whether other attention operations are replaced. Table 3 already specifies 96 input images, and Appendix B.2 states that inference uses a single H20 at 512 x 288 resolution. The remaining sequence configurations and outputs included in the timings for Tables 4-5 should be clarified.

Questions for authors
  1. How does the depth-consistency loss handle independently moving objects with nonzero depth-direction motion? Does the validity mask explicitly reject these points, or is an additional motion compensation applied? A static/dynamic-region breakdown would help establish the effect of this approximation.
  2. Were the baselines retrained or fine-tuned using the same training data as DriveWeave, including the six million self-collected frames? If the training data differ, could a matched-data comparison or an ablation excluding the self-collected corpus help separate methodological gains from data effects?
  3. Please specify the attention-layer schedule and synchronization frequency, and explain the latency reduction in Table 5. Beyond the configuration already stated for Table 3, what frame/view counts and prediction outputs are included in the timings for Tables 4-5?
3 comments

We sincerely thank the reviewer for the thorough, insightful, and constructive feedback. Below, we address each concern regarding the Flow-guided Depth Consistency (FDC) formulation on dynamic objects, baseline training comparability, and the timing/latency breakdown across Tables 3, 4, and 5.

Response to Question 1: Handling of Independently Moving Objects in FDC

Reviewer Concern: The depth-consistency constraint needs a clearer justification for independently moving objects. Equations (1)–(2) use optical flow to establish correspondences but transform target points using only ego-motion. Without additional masking or motion compensation, the loss may penalize correct depth changes on moving objects.

Our Response: We thank the reviewer for pointing out this potential ambiguity in our formulation. To clarify, FDC does not rely on rigid ego-motion warping alone to establish target depth coordinates; rather, it is anchored entirely by the estimated optical flow field Ft→t+1\mathbf{F}_{t \to t+1}.

  • Non-Rigid Correspondence via Optical Flow: The optical flow field Ft→t+1\mathbf{F}_{t \to t+1} maps pixel xt\mathbf{x}_{t} in frame tt to its actual corresponding location xt+1=xt+F(xt)\mathbf{x}_{t+1} = \mathbf{x}_{t} + \mathbf{F}(\mathbf{x}_{t}) in frame t+1t+1. For an independently moving object (e.g., an oncoming car), the flow vector captures the true 2D image-plane displacement resulting from both camera ego-motion and the object's independent velocity.
  • Avoiding Erroneous Penalties: Because correspondences are established directly through non-rigid optical flow rather than a rigid depth-pose back-projection, the temporal consistency constraint evaluates features and depth along the object's actual trajectory in the scene. Thus, FDC does not penalize legitimate depth changes of moving objects.
  • Role of the Validity Mask: Our validity mask filters out occlusions, sky regions, and out-of-bounds coordinates, serving as a reliability gate rather than an explicit semantic dynamic-object filter. FDC acts as a soft geometric regularization objective jointly optimized with per-frame dense LiDAR/pseudo-depth supervision.
  • Empirical Validation: As demonstrated in our qualitative visualizations (Fig.~4) and quantitative metrics (Table 6, showing consistent improvements in Temporal Absolute Error, TAE), DriveWeave maintains robust temporal consistency and sharp depth boundaries even in highly dynamic traffic scenarios.

Response to Question 2: Training Data Comparability and Dataset Ablation

Reviewer Concern: Training-data comparability with the baselines is unclear... If training data differ across methods, the cross-method gains may reflect both methodological and data differences.

Our Response: We appreciate this critical observation regarding data scaling versus architectural innovation. We evaluated DriveWeave under two distinct settings in Table R1 and Table R2 below (Pandase and Argoverse2 datasets are used for zero-shot evaluations):

  1. Model trained exclusively on public benchmarks (without our self-collected data).
  2. Model trained with the full self-collected dense LiDAR corpus.
  • Baseline Evaluation Protocol: The competing SOTA foundation models (e.g., DA3, Metric-anything) are evaluated using their official, publicly released pre-trained checkpoints. Since these models are not open-sourced for retraining on our proprietary dataset, direct retraining of all baselines on our 6-million-frame corpus was computationally prohibitive.
  • Isolating Architectural Gains via Controlled Ablation: To rigorously separate architectural improvements from dataset scale, we conducted a controlled ablation on DriveWeave itself. We evaluated our model when trained exclusively on standard public benchmarks (omitting the 6M self-collected LiDAR corpus) versus our full pipeline:

Key Takeaways from the Ablation:

  1. Dominance of Architecture: Even when trained strictly on the same public datasets as prior art, DriveWeave outperforms existing methods by a wide margin, confirming that our core innovations (4D-SyncAttn, FDC, and FRC) are the primary drivers of performance gains.
  2. Complementary Role of Data Scaling: Incorporating our 6M self-collected dense LiDAR corpus further boosts generalization, sensor-noise robustness, and long-horizon consistency. We have added this ablation and discussion to the revised experiment section.

Table R1. Ablation Study on Data Scaling. (Nuscenes [left half] / Waymo [right half]: Rel$\downarrow$, $\delta_1$$\uparrow,Acc, Acc\downarrow,AUC3, AUC3\uparrow,AUC30, AUC30\uparrow$)

Data ScalingRelδ1\delta_1AccAUC3AUC30Relδ1\delta_1AccAUC3AUC30
w/o self-collected10.288.10.5383.4286.829.890.70.5480.7783.42
full data10.088.50.5382.1986.679.690.20.5281.2483.81

Table R2. Comparison of Zero-shot Generalization Capability. (Pandaset [left half] / Argoverse2 [right half]: Rel↓\downarrow, \delta_1$$\uparrow\, Acc$\downarrow$, AUC3$\uparrow$, AUC30$\uparrow$)

MethodRelδ1\delta_1AccAUC3AUC30Relδ1\delta_1AccAUC3AUC30
DVGT [1]32.850.12.4078.2586.8453.234.66.4264.0584.15
Metric-anything [2]24.776.41.26--27.469.41.24--
w/o self-collected18.784.81.1385.3490.7719.180.40.9574.4687.85
full data16.486.61.0887.6991.2118.381.70.8881.2588.37

[1] Zuo et al. Dvgt: Driving visual geometry transformer. CVPR 2026.

[2] Ma et al. Metric-anything: Scaling metric depth pretraining with noisy heterogeneous sources. ECCV 2026.

Response to Question 3: Attention Schedule, Latency, and Timing Breakdown (Tables 3–5)

Reviewer Concern: Please specify the attention-layer schedule and synchronization frequency, and explain the latency reduction in Table 5. What frame/view counts and prediction outputs are included in the timings for Tables 4–5?

Our Response: All inference benchmarks were conducted on a single NVIDIA H20 GPU with an input resolution of 512×288512 \times 288 (as stated in Appendix B.2). We clarify the exact pipeline scope and tensor configurations for Tables 3, 4, and 5 below:

  1. Table 3 (Point Cloud Processing Time): Evaluate the time of feed-forward for all point clouds in a 16 × 6 image model, including feed-forward + point cloud processing (depth/point head to obtain the current local point cloud coordinate system, and then convert it to ego-first coordinate system).
  2. Table 4 (End-to-End Rendering Latency): Evaluates the full pipeline based on a standard 6-frame surround-view input, comprising backbone feed-forward + depth/point-cloud head inference + 3D Gaussian Splatting rendering.
  3. Table 5 (Core Backbone Latency): Evaluates strictly the computational overhead of the feature extraction and spatiotemporal attention blocks (excluding task-specific post-processing heads and rendering loops).

We sincerely thank the reviewer for their valuable time, insightful feedback, and constructive engagement. We remain readily available to address any further questions or provide additional clarifications during the discussion phase, and we look forward to your follow-up thoughts.

Thanks to the author's reply, the author has resolved most of my concerns. I will adjust my score to 3, but similarly, my confidence is 2 because I haven't conducted research directly related to this field.

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

DriveWeave jointly predicts metric depth, optical flow, ego-motion, and 3D Gaussians from calibrated surround-view videos. The model combines factorized spatial and temporal attention with synchronized register tokens, and uses flow-guided losses to encourage consistency between depth predictions and rendered images across frames. It is trained on a mixture of synthetic data, public driving datasets, and a large self-collected dataset. The paper reports improvements across depth estimation, pose estimation, point-cloud reconstruction, and rendering, with ablations examining the main components.

Critical strengths

The integration of geometry, motion, and rendering in a single feed-forward model is useful and well motivated. The paper reports strong results across several driving benchmarks, and the consistency-loss ablations provide evidence that coupling these tasks improves both geometry and rendering.

Critical weaknesses

The treatment of moving objects in FDC is unclear, since the formulation compensates only for ego-motion. Without additional masking or motion compensation, the loss may penalize correct depth changes on moving objects. The contribution of the large self-collected dataset is also not isolated through ablation, making it difficult to assess how much of the improvement comes from the proposed method versus additional training data.

Questions for authors
  1. As written, FDC accounts only for ego-motion and does not generally hold for independently moving objects. Are these objects masked out, or is their motion compensated for?
  2. The self-collected dataset is substantial, but its composition and contribution are not sufficiently documented. Could the authors provide more detailed statistics and an ablation without this dataset?
2 comments

We thank the reviewer for the insightful and constructive feedback. Below, we provide detailed clarifications and additional experimental results to address your concerns regarding the Flow-guided Depth Consistency (FDC) on moving objects and the isolation of our self-collected dataset.

Response to Question 1: Handling of Independently Moving Objects in FDC

Reviewer Concern: The treatment of moving objects in FDC is unclear, since the formulation compensates only for ego-motion. Without additional masking or motion compensation, the loss may penalize correct depth changes on moving objects.

Our Response: We apologize for the lack of clarity in our initial manuscript. FDC does not rely solely on ego-motion rigid warping; rather, it is guided by optical flow, which captures the total pixel displacement between consecutive frames (incorporating both camera ego-motion and independent object motion).

  • Why it does not penalize moving objects: Traditional ego-motion-only temporal consistency assumes a static scene and warps pixels purely using camera pose and depth, which indeed fails to penalize correct depth changes on moving cars/pedestrians. In contrast, our FDC utilizes the estimated optical flow field Ft→t+1\mathbf{F}_{t \to t+1} to establish true pixel-to-pixel correspondences in the image plane.
  • The Mechanism: For an independently moving object, the optical flow vector points directly to the object's actual displaced position in the next frame. Therefore, matching features or depth along the flow trajectory tracks the physical motion of the object rather than assuming a static background constraint. As a result, FDC avoids introducing erroneous penalties on moving objects without requiring heuristic semantic masks or separate dynamic motion branches.
  • Empirical Evidence: As demonstrated in qualitative visualizations (Fig. 4), DriveWeave maintains sharp depth consistency across dynamic traffic scenes. Furthermore, the consistent improvements in Temporal Absolute Error (TAE) across dynamic driving benchmarks (Table 6) validate that FDC effectively regularizes temporal stability under realistic driving conditions.

Response to Question 2: Statistics and Ablation of the Self-Collected Dataset

Reviewer Concern: The contribution of the large self-collected dataset is not isolated through ablation, making it difficult to assess how much of the improvement comes from the proposed method versus additional training data.

Our Response: To isolate our architectural contributions from dataset scaling effects, we conducted rigorous ablation experiments. Specifically, we evaluated DriveWeave under two distinct settings in Table R1 and Table R2 below (Pandase and Argoverse2 datasets are used for cross-dataset evaluations without any target-domain fine-tuning or re-training):

  1. Model trained exclusively on public benchmarks (without our self-collected data).
  2. Model trained with the full self-collected dense LiDAR corpus.

Analysis of Results:

  • Methodological Superiority: As can be seen in the Tables below, even when trained strictly on standard public datasets (without our self-collected corpus), DriveWeave outperforms existing state-of-the-art methods by a significant margin. This confirms that our core architectural innovations (4D-SyncAttn and flow-guided consistency) are the primary drivers of performance gains.
  • Data Scaling Benefits: As shown in Table R2 below, the models are tested on unseen datasets in training: Pandaset and Argoverse2. The model trained with the full data achieves the best performance, which implies that incorporating our self-collected 6M-frame dense LiDAR corpus further boosts zero-shot generalization, depth coverage in complex conditions, and long-horizon robustness. The dataset acts as a powerful complement that unlocks the full capacity of our unified foundation model without altering its underlying mechanics.

Table R1. Ablation study on data scaling. (Nuscenes [left half] / Waymo [right half]: Rel$\downarrow$, $\delta_1$$\uparrow,Acc, Acc\downarrow,AUC3, AUC3\uparrow,AUC30, AUC30\uparrow$)

Data ScalingRelδ1\delta_1AccAUC3AUC30Relδ1\delta_1AccAUC3AUC30
w/o self-collected10.288.10.5383.4286.829.890.70.5480.7783.42
full data10.088.50.5382.1986.679.690.20.5281.2483.81

Table R2. Comparison of Zero-shot Generalization Capability. (Pandaset [left half] / Argoverse2 [right half]: Rel↓\downarrow, \delta_1$$\uparrow\, Acc$\downarrow$, AUC3$\uparrow$, AUC30$\uparrow$)

MethodRelδ1\delta_1AccAUC3AUC30Relδ1\delta_1AccAUC3AUC30
DVGT [1]32.850.12.4078.2586.8453.234.66.4264.0584.15
Metric-anything [2]24.776.41.26--27.469.41.24--
w/o self-collected18.784.81.1385.3490.7719.180.40.9574.4687.85
full data16.486.61.0887.6991.2118.381.70.8881.2588.37

[1] Sicheng Zuo, Zixun Xie, Wenzhao Zheng, Shaoqing Xu, Fang Li, Shengyin Jiang, Long Chen, Zhi-Xin Yang, and Jiwen Lu. Dvgt: Driving visual geometry transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.

[2] Baorui Ma, Jiahui Yang, Donglin Di, Xuancheng Zhang, Jianxun Cui, Hao Li, Xie Yan, and Wei Chen. Metric-anything: Scaling metric depth pretraining with noisy heterogeneous sources. In Proceedings of the European Conference on Computer Vision (ECCV), 2026.

We thank the reviewer once again for their valuable time, constructive feedback, and recognition of our work's strengths. We remain entirely at your disposal should any further questions arise during the discussion phase.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.