acceptodds
Under review as a conference paper at ICLR 2027

ForceStereo: Observation Assimilation through State Evolution for Stereo Matching

Abstract

Modern stereo matching methods achieve strong overall accuracy by combining stereo matching cues, contextual features, and monocular depth priors. However, they rarely assess how reliable different image regions are in open world scenes, so overall metrics and visual results can obscure large errors in regions with unreliable observations. These hidden errors may later surface as noise or artifacts in downstream 3D pipelines. We propose ForceStereo, a depth state inference framework based on observation calibration that reformulates disparity estimation as state evolution guided by reliability. ForceStereo first uses scene-level context to calibrate ambiguous local disparity hypotheses before state formation, so the model does not commit too early to unreliable matches. It then interprets residuals, hypothesis ambiguity, and structural information as reliability conditions, and uses a Neural ODE to regulate their influence throughout disparity evolution. This inference process controls how unreliable observations affect the evolving disparity state instead of directly turning them into disparity updates, improving robustness and generalization in open world scenes. Extensive experiments establish state-of-the-art performance, including up to 16.0% error reduction on the ETH3D leaderboards and up to 25.3% zero-shot error reduction over prior best results on challenging datasets. ForceStereo ranks 1st on multiple key metrics in both the KITTI 2012 and KITTI 2015 stereo benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
stereo matching, observation assimilation, state-space inference, reliability modeling, continuous dynamics, zero-shot generalization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RYCEOpenReview ↗
Rating
3(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2(−1)
Read review
Paper summary

The paper presents ForceStereo, a stereo matching framework that treats disparity estimation as reliability-guided state evolution. The method first uses scene-level context to calibrate local disparity hypotheses through an Overview-Conditioned Hypothesis Volume (OCHV). It then constructs reliability conditions from state-dependent residuals and structural cues, followed by a Neural ODE-based Residual Continuous Evolution (RCE) module to update the disparity state. The method achieves strong results on several benchmarks and zero-shot evaluations. The ablation study validates the effectiveness of the proposed modules.

Critical strengths
  1. The paper provides a meaningful perspective on stereo matching by treating observation reliability as part of the inference process, rather than as an auxiliary confidence prediction.
  2. The experimental results are strong and consistent across standard benchmarks and zero-shot settings. The ablation study is also comprehensive and helps clarify the contribution of the proposed components.
  3. The paper is clearly written and logically organized, making the proposed method and experimental findings easy to follow.
Critical weaknesses
  1. The practical distinction between the Neural ODE formulation and a carefully designed iterative update is not completely clear. The method is implemented with three fixed-step Euler evaluations, although the paper presents it as continuous state evolution. The matched-GRU comparison is helpful, but more discussion would be useful on whether the improvement comes from the continuous formulation itself or from the additional state-dependent conditioning and trajectory supervision.

  2. The accuracy improvements are well demonstrated, but the efficiency evaluation is less comprehensive. The paper should report the runtime and memory cost on one or two representative datasets and under different image resolutions. It would also be useful to separate the costs of feature extraction, reliability computation, and RCE updates, and to discuss the resulting accuracy–efficiency trade-off.

Questions for authors
  1. How sensitive are the results to the number of RCE solver steps and the integration interval? Does increasing the number of steps provide further improvement or mainly increase computation?
  2. How are the model-independent unreliability scores and their thresholds defined in the zero-shot analysis?
  3. What is the complete inference cost of ForceStereo compared with BridgeDepth and the other baselines, including memory usage and the additional feature warping and residual computation?
8 comments
AuthorsOpenReview ↗

General Response

We thank the reviewer for the constructive feedback and recognition of our core contribution. We appreciate the thoughtful questions on continuous evolution, computational efficiency, and evaluation methodology.

We have also revised the method description to improve clarity and readability. Additional comparison and ablation experiments have been included, particularly to examine the role of monocular priors. All revised text, figures, and tables are highlighted in blue for easy identification.

Below, we address each concern separately with experimental evidence and additional analyses.

AuthorsOpenReview ↗

W1. Continuous Evolution and Discrete Refinement

W1. Continuous Evolution and Discrete Refinement

Where the gains come from. The key benefit of RCE lies in how state-dependent observations are converted into geometric changes. As disparity evolves, RP regenerates residuals and matching geometry. RCE maps these updated conditions to ambiguity-bounded local velocities and integrates them to refine the geometry state. This couples observation feedback with a rate-based evolution rule.

Continuous evolution vs. discrete GRU refinement. RCE defines this rule through a shared velocity field over a continuous evolution coordinate. Three fixed-step Euler evaluations numerically approximate its integration. A conditioned GRU also receives the updated RP, but directly predicts finite gated state transitions. Its updates depend on learned gates and state features, without explicitly linking local geometric rates to an integration step size. RCE explicitly accumulates regulated local velocities through integration. The key distinction is how the same observations govern state changes, not access to additional conditioning or supervision.

1. Controlled comparison with discrete refinement. To test the benefit of this velocity-field parameterization, we compare RCE with a matched GRU. Both use the same RP regeneration, ambiguity gating, trajectory supervision, and three-step update budget (Table 6; Appendix Table 13). Beyond standard benchmarks, we also evaluate both models on the most unreliable 10% of Middlebury pixels, identified by our model-independent unreliability score.

MethodMiddlebury BP-2Unreliable 10% BP-2ETH3D BP-1
Matched GRU (3 steps)7.7311.161.10
RCE (3 steps)7.6210.030.97

RCE achieves lower errors under the same conditioning and supervision. In highly unreliable regions, it achieves a 10.1% relative reduction in BP-2 (11.16% to 10.03%). These results support an additional benefit from RCE's velocity-field update rule beyond state-dependent conditioning and trajectory supervision.

2. Contribution of state-dependent observation feedback. We separately examine the benefit of dynamically updating observation conditions within RCE (Appendix Table 12).

RP ConfigurationMiddlebury BP-2Unreliable 10% BP-2
Fixed RP7.9111.24
Dynamic Residual7.7910.90
Dynamic Geometry7.8210.68
Full Dynamic RP7.6210.03

Updating both residuals and matching geometry reduces Middlebury BP-2 from 7.91% to 7.62%, with a larger relative improvement in unreliable regions (11.24% to 10.03%). This confirms that reinterpreting observations as the geometry state evolves also contributes to performance.

Together, these experiments support two complementary sources of improvement. State-dependent observation feedback improves the conditions guiding geometric refinement. RCE's velocity-field parameterization provides an additional advantage over the matched GRU under the same conditioning and supervision. Their combination enables reliability-guided geometric evolution.

References

[1] Neural Ordinary Differential Equations. Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, David K. Duvenaud.

[2] RAFT-Stereo: Multilevel Recurrent Field Transforms for Stereo Matching. Lahav Lipson, Zachary Teed, Jia Deng.

AuthorsOpenReview ↗

W2 and Q3. Inference Efficiency and Computational Cost

W2 and Q3. Inference Efficiency and Computational Cost

ForceStereo improves stereo accuracy with a 16 ms increase in end-to-end inference time over BridgeDepth on Scene Flow. Latency increases from 75 ms to 91 ms. To quantify this trade-off, we additionally profile inference latency, peak GPU memory, and component costs on an NVIDIA L20 GPU.

End-to-end runtime and memory. We measure both methods at three input resolutions.

Input ResolutionBridgeDepth LatencyForceStereo LatencyBridgeDepth Peak MemoryForceStereo Peak Memory
496 × 72068.34 ms84.85 ms1.640 GiB1.871 GiB
992 × 1440152.14 ms184.59 ms2.633 GiB3.547 GiB
1984 × 2880887.21 ms959.75 ms6.627 GiB8.263 GiB

The additional latency is 16.51, 32.45, and 72.54 ms, respectively. As input resolution increases, ForceStereo's relative inference latency overhead over BridgeDepth decreases from 24.2% to 8.2%. The corresponding peak memory measurements are reported in the table.

Component-wise profiling. We measure the main operations separately on the L20 GPU.

ComponentLatency
Feature Extraction (DAv2 + Stereo)25.99 ms
Proposal / Hypothesis / State15.45 ms
OCHV7.95 ms
Latent Alignment24.21 ms
RP (3 steps)5.91 ms
RCE excluding RP (3 steps)11.64 ms

Across three Euler steps, RP takes 5.91 ms, while RCE excluding RP takes 11.64 ms. The latter includes dynamic geometry computation and Euler updates. These are individually profiled component timings, and end-to-end inference is measured separately.

Warping and residual costs. We further decompose the RP computation across three Euler steps.

RP Operation (3-step total)Latency
Image Warping0.54 ms
Feature Warping + Normalization0.65 ms
Residual + Frequency Computation0.71 ms
Reliability Features + Fusion4.01 ms
Total RP5.91 ms

Image warping, feature warping with normalization, and residual/frequency computation together account for 1.90 ms across three steps. Reliability feature construction and fusion account for the remaining 4.01 ms. These measurements directly quantify the observation-processing costs of RP.

Accuracy–efficiency trade-off. The additional computation is accompanied by improved accuracy over BridgeDepth (Table 4).

BenchmarkBridgeDepthForceStereoRelative Error Reduction
ETH3D BP-11.260.9723.02%
Middlebury BP-27.847.622.81%
DrivingStereo BP-34.283.5916.12%

ForceStereo adds 1.01M parameters (0.29%). Table 7 also shows comparable absolute latency increases when applying the framework to other backbones: IGEV-Stereo increases from 375 to 389 ms (+14 ms), and NMRF from 42 to 58 ms (+16 ms).

ModelMiddlebury BP-2ETH3D BP-1Latency
IGEV-Stereo9.954.05375 ms
Force-IGEV8.983.32389 ms
NMRF8.533.7142 ms
Force-NMRF8.012.5658 ms

Overall, these measurements quantify the computational cost of dynamic observation feedback and reliability-guided state evolution. With 16 ms additional end-to-end latency on Scene Flow, ForceStereo achieves 23.0% and 16.1% relative error reductions on ETH3D and DrivingStereo, respectively. These results establish a clear accuracy–efficiency trade-off for the proposed refinement mechanism.

We have organized the implementations of OCHV, RP, and RCE, along with the profiling scripts, in a maintained code repository. We plan to release it promptly upon acceptance to support reproducibility and further research.

References

[1] BridgeDepth: Bridging Monocular and Stereo Reasoning with Latent Alignment. Tongfan Guan, Jiaxin Guo, Chen Wang, Yun-Hui Liu.

[2] Iterative Geometry Encoding Volume for Stereo Matching. Gangwei Xu, Xianqi Wang, Xiaohuan Ding, Xin Yang.

[3] Neural Markov Random Field for Stereo Matching. Tongfan Guan, Chen Wang, Yun-Hui Liu.

Reviewer_f2RVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper focuses on improving stereo prediction on ambiguous regions of an image, e.g. textureless surface, occluded in one view. It proposes ForceStereo, which iteratively updates stereo prediction in an reliability-aware fashion. In each iteration of update, the model first calculates features of how reliable the correspondence is, which involves ambiguity cue, edge cue, dynamic geometry, etc.. Then, the model updates prediction based on these reliability features.

ForceStereo is evaluated on common stereo benchmarks, and is better than almost all baselines under standard metrics. Qualitatively, ForceStereo shows better performance on occluded regions (regions visible in only 1 view).

Critical strengths
  • The method is well motivated. It keeps multiple hypotheses, and updates predictions based on reliability features. Moreover, the design of reliability features makes sense. Each factor, e.g. dynamic geometry, makes correspondence less reliable. This design is also proved useful in ablation study (Table 5).
  • Authors also show that their methodology can be applied to other models, e.g. IGEV-Stereo and NMRF, though it lacks more discussion on how to generalize.
Critical weaknesses
  • Quantitative results are lacking. Despite comparing with many baselines, there is no comparison with some state-of-the-art stereo models, including waft-stereo [1] and foundation stereo [2] in all 3 quantitative results (Table 1,2,3).

  • The writing makes the paper hard to understand. Too many variables and notations are defined in Sec 3, and some lack clear definition. For example, in Eq. (7), the paper only mentions δr\delta_r and δw\delta_w perform residual interpretation and adaptive gating, but what are these two notations? Are they a variable or a neural network layer? Moreover, some figures are hard to follow, e.g. no caption explains what heatmap represents in Fig. 4..

[1] WAFT-Stereo: Warping-Alone Field Transforms for Stereo Matching. Yihan Wang, Jia Deng

[2] FoundationStereo: Zero-Shot Stereo Matching. Bowen Wen, Matthew Trepte, Joseph Aribido, Jan Kautz, Orazio Gallo, Stan Birchfield

Questions for authors
  • How does the performance of ForceStereo compare with WAFT-Stereo and FoundationStereo?

  • Can authors explain what is shown in heatmap of Fig. 4, and what does warmer or colder color mean?

  • Can authors elaborate on “Module Universality” (Line 469-473)? How can ForceStereo be integrated to other stereo methods (e.g. what part of model is switched)?

  • ForceStereo contains a monocular depth estimation module (DAv2). How does the quality of monocular prior affects ForceStereo performance, especially on unreliable region? Have authors tried different monocular depth backbones?

5 comments
AuthorsOpenReview ↗

General Response

We thank the reviewer for the careful evaluation and constructive feedback. We appreciate the thoughtful questions regarding recent stereo comparisons, method clarity and visualization, module universality, and the influence of monocular depth priors.

We have made revisions to improve the presentation and readability of the manuscript. We refined the method descriptions, streamlined the mathematical notation by reducing the number of symbols in the equations, and clarified the definitions and roles of the remaining variables. These changes make the formulations easier to follow while preserving the technical details. We also improved the relevant figure explanations and captions. All revised text, figures, and tables are highlighted in blue for easy identification.

We further strengthened the experimental evaluation with additional comparisons against recent stereo methods, analyses of module integration across different backbones, and a monocular backbone sensitivity study across challenging datasets.

Below, we address each concern separately, providing the corresponding explanations, revisions, and experimental evidence. We hope these improvements make the contributions of ForceStereo clearer and further demonstrate the value of reliability-guided stereo inference.

AuthorsOpenReview ↗

W1/Q1. Expanded SOTA Comparisons and Training Scale

W1/Q1. Expanded SOTA Comparisons and Training Scale

We added WAFT-Stereo and MonSter++ to the official benchmark comparison (revised Table 2), and WAFT-Stereo and FoundationStereo to the expanded-pretraining zero-shot comparison (Table 3). Full metrics for the requested methods follow.

Evaluation coverage. Table 1 uses Scene Flow-only training and evaluation, for which neither WAFT-Stereo nor FoundationStereo provides a public matching checkpoint and test metrics. Table 2 uses official leaderboard submissions; FoundationStereo has no corresponding KITTI 2012/2015 entry under its method name. We therefore compare the available results in Table 3 rather than mixing evaluation protocols.

Training Scale and Reproducibility. For official benchmarks, we added the FoundationStereo Dataset (FSD) to pretraining before benchmark-specific fine-tuning (Sec. 4). Zero-shot tests use no target-dataset fine-tuning. Reported pretraining scales:

  • ForceStereo: 300K steps × batch 8 = 2.4M sample exposures.
  • WAFT-Stereo: 400K × 32 = 12.8M.
  • FoundationStereo: 200K × 128 = 25.6M.

WAFT-Stereo and FoundationStereo use 5.3× and 10.7× as many sample exposures as ForceStereo, respectively. ForceStereo nevertheless achieves competitive accuracy with a clearly specified, smaller pretraining budget.

Complete Official Benchmark Comparison (Table 2). With FSD-expanded pretraining and updated weights, ForceStereo retains its previously reported KITTI 2012 results at the reported precision. KITTI 2015 D1-all Noc improves from 1.28% to 1.20%, and D1-all All from 1.35% to 1.26%. Each cell below reports Noc / All; lower is better, and best values (including ties) are bold. Values are percentages except ETH3D RMSE (px).

BenchmarkMetricWAFT-StereoForceStereo
KITTI 2012 ReflectiveOut-25.31 / 6.375.16 / 5.90
KITTI 2012 ReflectiveOut-32.74 / 3.392.49 / 2.92
KITTI 2012 ReflectiveOut-41.90 / 2.371.61 / 1.90
KITTI 2012Out-21.18 / 1.471.24 / 1.56
KITTI 2012Out-30.77 / 0.950.77 / 0.99
KITTI 2012Out-40.59 / 0.730.59 / 0.73
KITTI 2015D1-bg0.98 / 1.051.03 / 1.09
KITTI 2015D1-fg2.34 / 2.472.04 / 2.09
KITTI 2015D1-all1.21 / 1.281.20 / 1.26
ETH3DBad 1.00.28 / 0.310.32 / 0.42
ETH3DRMSE0.20 / 0.210.19 / 0.24

ForceStereo improves all six KITTI 2012 Reflective metrics over WAFT-Stereo and four of six KITTI 2015 metrics. In particular, KITTI 2015 D1-fg All decreases from 2.47% to 2.09% (15.4% relative). These gains align with our focus on unreliable matching near reflections and foreground boundaries, where appearance inconsistency does not necessarily indicate incorrect disparity. Official KITTI leaderboard screenshots are provided in Appendix Figure 11.

Complete Expanded-Pretraining Zero-Shot Comparison (Table 3). Each cell gives EPE (px) / BP (%): BP-3 for KITTI and weather, BP-2 for Middlebury, BP-1 for ETH3D. Best values among the three methods are bold.

DatasetFoundationStereoWAFT-StereoForceStereo
KITTI 20120.68 / 3.000.69 / 3.110.68 / 3.09
KITTI 20150.87 / 3.100.87 / 3.130.86 / 3.27
Middlebury0.66 / 3.110.80 / 3.070.78 / 3.87
ETH3D0.17 / 0.670.15 / 0.530.17 / 1.04
Cloudy0.91 / 2.940.95 / 2.970.91 / 2.78
Foggy0.92 / 2.521.15 / 5.341.10 / 5.03
Rainy1.44 / 11.261.21 / 5.431.12 / 4.97
Sunny0.94 / 3.500.98 / 3.240.93 / 3.01

FoundationStereo and WAFT-Stereo lead on some Middlebury and ETH3D metrics. ForceStereo matches FoundationStereo on KITTI 2012 EPE, achieves the lowest KITTI 2015 EPE, and leads BP-3 on Cloudy, Rainy, and Sunny, with far fewer sample exposures. On Rainy, BP-3 decreases from WAFT-Stereo's 5.43% to 4.97%.

Relation to Our Contribution. OCHV calibrates ambiguous hypotheses before refinement. RP updates residual and geometric cues as disparity evolves, and RCE uses them to regulate state changes. This allows observation reliability to guide geometric refinement as matching evidence changes. Figure 6 shows slower error growth with increasing observation unreliability, and Tables 4–6 support the individual modules. These results link competitive performance in challenging settings to reliability-guided inference.

References

[1] WAFT-Stereo: Warping-Alone Field Transforms for Stereo Matching. Yihan Wang, Jia Deng. ECCV 2026.

[2] MonSter++: Unified Stereo Matching, Multi-View Stereo, and Real-Time Stereo With Monodepth Priors. Junda Cheng et al. IEEE TPAMI, 2026.

[3] FoundationStereo: Zero-Shot Stereo Matching. Bowen Wen et al. CVPR 2025.

AuthorsOpenReview ↗

W2/Q2. Clarification of Method Notations and Figure 4

W2/Q2. Clarification of Method Notations and Figure 4

We thank the reviewer for pointing out the unclear notation in Sec. 3 and the missing explanation of Figure 4. We agree that these descriptions should be more explicit. We have revised Secs. 3.2–3.4 to clarify the computation and notation, and updated the Figure 4 caption to explain the visualization.

1. Notation and Method Clarity (Eq. 7).

In Eq. (7), δrρ\delta_r^\rho and δω\delta_\omega denote learnable neural network modules, not state variables. Specifically, δrlow\delta_r^{\mathrm{low}} and δrhigh\delta_r^{\mathrm{high}} are two separate convolutional branches. They encode the low- and high-frequency residuals together with hypothesis ambiguity UU and edge information EE, producing residual features BtlowB_t^{\mathrm{low}} and BthighB_t^{\mathrm{high}}.

The gating network δω\delta_\omega then processes these features with UU and EE. Its sigmoid output Ωt\Omega_t provides spatial weights for adaptively combining the two residual components. The fused features are concatenated with aligned context CalignC_{\mathrm{align}} and current matching geometry GtG_t to form the reliability condition PtP_t.

We have clarified these definitions and their computation in revised Sec. 3.3. We also refined the surrounding descriptions to better explain how observation encoding and hypothesis calibration support subsequent geometry refinement. During state evolution, RP recomputes residuals and matching geometry from the current disparity, while hypothesis ambiguity UU and edge information EE remain fixed. RCE then uses the updated reliability condition to guide the evolution of the geometry state.

2. Interpretation of Figure 4.

Figure 4 visualizes disparity estimates, not confidence, reliability, or error maps. Warmer colors indicate larger disparities, while cooler colors indicate smaller disparities. We have explicitly added this explanation to the revised figure caption.

The columns show the left input image and disparity estimates obtained from LHV, OHV, and OCHV, before subsequent processing. LHV uses local stereo matching evidence, while OHV introduces scene-level context. OCHV combines both representations over the same retained disparity proposals, preserving local matching details while improving the interpretation of ambiguous candidates.

We also strengthened the explanation in Sec. 3.2 and Appendix C.2. To directly demonstrate the calibration effect, we reproduce the results of Appendix Tables 9–10 below.

Hypothesis Calibration (Appendix Table 9). All representations are evaluated on the same retained disparity proposals. Higher selection accuracy and margin are better; lower NLL, entropy, and ambiguity are better.

RepresentationSel. Acc. (%) ↑NLL ↓HKH_K ↓Δ2\Delta_2 ↑UU ↓
LHV78.60.5170.6410.3360.653
OHV77.40.5340.6320.3490.642
OCHV83.30.4480.5680.4310.568

Calibration across Ambiguity Levels (Appendix Table 10). Pixels are grouped into three equal-sized strata using pre-calibration LHV ambiguity. The same groups are used for both representations.

Ambiguity StratumLHV (%)OCHV (%)Gain (pp)
T1 (0–33.3%)93.494.3+0.9
T2 (33.3–66.7%)81.785.4+3.7
T3 (66.7–100%)60.770.2+9.5

OCHV improves overall candidate-selection accuracy from 78.6% to 83.3% while reducing hypothesis ambiguity. Its gain increases with initial ambiguity, reaching 9.5 percentage points in the highest-ambiguity group. These results support the complementary roles of local matching evidence and scene context illustrated in Figure 4.

We appreciate the reviewer's suggestions, which helped us improve the clarity of the method descriptions and visualizations.

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

The paper presents the ForceStereo, a framework that reformulates disparity estimation as state evolution guided by reliability. The framework achieves good performance on KITTI 2012 and KITTI 2015.

Critical strengths
  1. The framework achieves good performance on public benchmarks KITTI 2012 and KITTI 2015.
  2. It's interesting to incorporate the information of matching error into network inputs, and use neural ODE to update the disparity and matching error iteratively.
Critical weaknesses
  1. There is prior work, WAFT-Stereo [1], that achieves better performance on KITTI 2012 and KITTI 2015 benchmark, which is ignored by the paper.
  2. The writing is difficult to understand. For example, equation (1) describes attention, but the "key" in the original attention is called "overview anchor" in your paper. Is there essential difference between the "key" and "anchor", or is the word "anchor" selected casually? Another example, how is the C_{align} computed? I didn't find explanation in the paper.
  3. Some details for experiments are missing. For example, what is detailed setting for the "Full w/o RP" in Table 4? Do you retrain the network or simply remove it only during test? And what is the input for the RCE after removing RP? There seems to be conflict between Table 4 and Table 5. For the Middlebury dataset, In Table 4, removing full RP gets 7.70 for BP-2. In Table 5, removing arbitrary component of RP gets higher error than 7.70 for BP-2 (except for edge cue).

[1] WAFT-Stereo: Warping-Alone Field Transforms for Stereo Matching. ECCV 2026.

Questions for authors
  1. Could you please explain the questions in the weakness?
5 comments
AuthorsOpenReview ↗

General Response

We thank the reviewer for the careful evaluation and constructive feedback. We appreciate the thoughtful questions regarding method comparisons, technical clarity, and RP ablation results.

In response, we have made substantial revisions to improve the presentation and readability of the manuscript. All revised text, figures, and tables are highlighted in blue for easy identification. We clarified key terminology, formulations, and implementation details. We also expanded comparisons with WAFT-Stereo and MonSter++, conducted additional zero-shot generalization experiments, and included official KITTI leaderboard screenshots in Appendix Figure 11. Furthermore, we clarified the ablation settings and strengthened the analysis of RP and dynamic observation feedback.

Below, we address W1–W3 separately, providing detailed explanations and supporting experimental evidence. We hope these revisions help clarify the contributions and empirical strengths of ForceStereo.

AuthorsOpenReview ↗

W1. Expanded SOTA Comparisons and Training Protocols

W1. Expanded SOTA Comparisons and Training Protocols

We added WAFT-Stereo and MonSter++ to the official benchmark comparison (Table 2), and expanded zero-shot evaluation to include FoundationStereo and WAFT-Stereo under broader synthetic pretraining (Table 3).

Expanded pretraining and training scale. We added the FoundationStereo Dataset (FSD) to our multi-dataset pretraining to broaden data coverage, although training budgets still differ. Benchmark-specific fine-tuning follows (Sec. 4). The updated model preserves the reported KITTI 2012 error rates at the shown precision and remains competitive on KITTI 2015 (Table 2).

WAFT-Stereo uses SynLarge, with approximately 7.7M pairs from 12 synthetic datasets (Sec. 4.1). Our pretraining uses 7 source datasets, including FSD:

Pretraining BudgetForceStereoWAFT-Stereo
Training steps300K400K
Global batch size832
Samples processed (steps × batch)2.4M12.8M
Source datasets712

ForceStereo uses 100K fewer steps and 5.3× fewer sample exposures than WAFT-Stereo.

Official benchmark comparison (revised Table 2). All reported metrics for six representative methods are shown. Lower is better; values are percentages except ETH3D RMSE (px). Bold, (2nd), and (3rd) mark the top three among listed methods; ties share rank.

Benchmark / MetricBridgeDepthMonSterPromptStereoWAFT-StereoMonSter++ForceStereo
KITTI 2012 Reflective
Out-2 Noc5.805.665.115.31 (3rd)5.685.16 (2nd)
Out-2 All6.856.816.17 (2nd)6.37 (3rd)6.665.90
Out-3 Noc2.912.752.54 (2nd)2.74 (3rd)2.912.49
Out-3 All3.483.38 (3rd)3.08 (2nd)3.393.452.92
Out-4 Noc1.881.73 (3rd)1.611.901.921.61
Out-4 All2.232.13 (3rd)1.91 (2nd)2.372.291.90
KITTI 2012 Standard
Out-2 Noc1.321.361.321.181.30 (3rd)1.24 (2nd)
Out-2 All1.65 (3rd)1.751.691.471.701.56 (2nd)
Out-3 Noc0.830.840.840.770.79 (3rd)0.77
Out-3 All1.03 (3rd)1.091.090.951.070.99 (2nd)
Out-4 Noc0.640.640.650.590.60 (3rd)0.59
Out-4 All0.77 (3rd)0.810.830.730.810.73
KITTI 2015
D1-bg Noc1.051.051.040.981.02 (2nd)1.03 (3rd)
D1-fg Noc2.62 (3rd)2.762.762.34 (2nd)2.672.04
D1-all Noc1.311.331.321.21 (2nd)1.29 (3rd)1.20
D1-bg All1.131.131.12 (3rd)1.051.12 (3rd)1.09 (2nd)
D1-fg All2.732.812.852.47 (2nd)2.65 (3rd)2.09
D1-all All1.401.411.411.28 (2nd)1.37 (3rd)1.26
ETH3D
Bad 1.0 Noc0.380.46–0.28 (2nd)0.250.32 (3rd)
Bad 1.0 All0.500.72–0.310.45 (3rd)0.42 (2nd)
RMSE Noc0.19 (2nd)0.20–0.200.180.19 (2nd)
RMSE All0.26 (3rd)0.47–0.210.340.24 (2nd)

Performance analysis. ForceStereo ranks first (including ties) on five of six KITTI 2012 Reflective metrics and second on the sixth. It outperforms WAFT-Stereo on all six. On KITTI 2015, it leads four of six metrics. D1-fg All improves from WAFT-Stereo's 2.47% to 2.09% (15.4% relative reduction), while D1-all All reaches 1.26% versus 1.28%. Official KITTI leaderboard screenshots are provided in Appendix Fig. 11.

These results support our focus on unreliable matching observations. OCHV calibrates disparity hypotheses, while RP updates reliability cues during RCE state evolution. Table 3 evaluates zero-shot generalization under distinct training protocols. Figure 6 shows slower error growth in unreliable regions on Foggy and Middlebury, using model-independent region definitions. Tables 4–6 further support the contributions of OCHV, RP, and RCE.

References

[1] WAFT-Stereo: Warping-Alone Field Transforms for Stereo Matching. Yihan Wang, Jia Deng.

[2] MonSter++: Unified Stereo Matching, Multi-View Stereo, and Real-Time Stereo With Monodepth Priors. Junda Cheng, Wenjing Liao, Zhipeng Cai, Longliang Liu, Gangwei Xu, Xianqi Wang, Yuzhou Wang, Zikang Yuan, Yong Deng, Jinliang Zang, Yangyang Shi, Jinhui Tang, Xin Yang.

AuthorsOpenReview ↗

W2. Clarification of Attention Terminology and Aligned Context

W2. Clarification of Attention Terminology and Aligned Context

1. Overview anchors and attention keys (Eq. 1).

We used "overview anchors" to describe pooled scene-overview tokens serving as spatial references for local stereo features, following prior anchor-token terminology (e.g., AnchorFormer, Shan et al., 2025). Their projections serve as standard keys in the Query–Key affinity computation. We appreciate the reviewer's observation that the distinction was unclear. We therefore adopted standard Query–Key terminology in revised Sec. 3.2 while preserving the original computation.

A lightweight CNN and adaptive pooling produce the overview OO. Local stereo features provide queries, and overview tokens provide keys:

K=\operatorname{Re}(W_kO),\quad A_v=\operatorname{Softmax}(Q_v^\top K).$$ **Overview-guided hypothesis calibration.** Rather than aggregating attention values, we use local-to-overview affinities $A_v$ to generate dynamic convolution kernels with two receptive fields: $$(W_v^S,W_v^L)=\operatorname{Split}(\operatorname{FFN}(A_v)).$$ The small kernels preserve fine matching details, while the large kernels capture broader scene context. Together, they produce overview-conditioned stereo features for constructing OHV. For each disparity candidate, OHV provides scene-conditioned matching evidence, while LHV retains original local matching support. We fuse OHV, LHV, and disparity encoding into OCHV (Eqs. 2–3). **This helps resolve ambiguous local hypotheses without discarding fine-grained evidence.** **2. Computation of $C_{\mathrm{align}}$ (Sec. 3.2 and Appendix A.1).** $C_{\mathrm{align}}$ is the aligned stereo context obtained after two stages of monocular–stereo feature alignment. We clarified its origin and role in revised Sec. 3.2, and detailed its computation in Appendix A.1 (Eqs. 12–14). Both stages follow BridgeDepth's bidirectional alignment. **Alignment operation.** Stereo features first read monocular structure through windowed cross-attention, followed by neural message passing (NMP). The updated stereo features then provide geometry to the monocular branch: $$\begin{gathered} \widetilde S_i = \Psi_{s\leftarrow m}(S_i,M_i),\\ S_{i+1} = \operatorname{NMP}(\widetilde S_i),\\ M_{i+1} = \Psi_{m\leftarrow s}(M_i,S_{i+1}). \end{gathered}$$ Here, $S_i$ and $M_i$ are stereo and monocular features, and $\Psi$ denotes windowed cross-attention. **Stage 1: Candidate-level alignment.** We extract stereo candidate features from OCHV and align them with structural features from the frozen monocular branch. This produces $S_{\mathrm{cand}}$, combining stereo matching evidence with monocular structural context. Candidate scores are normalized into a disparity hypothesis distribution, supporting disparity selection and ambiguity estimation $U$ (Appendix A.2). **Stage 2: Alignment after disparity selection.** We reconstruct stereo embeddings using the selected disparity and align them again with updated monocular features. The resulting stereo representation is $C_{\mathrm{align}}$. This provides structural context for the selected disparity instead of the full set of competing candidates. **Role in subsequent evolution.** $C_{\mathrm{align}}$ provides structural context for RP to interpret observation reliability during RCE. It also supports geometry initialization: a disparity head predicts local patch-wise corrections $\Delta D_{\mathrm{patch}}$ from $C_{\mathrm{align}}$. These corrections refine the selected coarse disparity: $$h(0)=\operatorname{ReLU}(D_0^{\mathrm{coarse}}+\Delta D_{\mathrm{patch}}).$$ RCE then evolves the explicit geometry state $h(t)$, and the final disparity is obtained directly from $h(1)$ (Appendix A.3). Thus, $C_{\mathrm{align}}$ links OCHV calibration with subsequent reliability-guided evolution. For **reproducibility**, we provided **anonymized code and execution logs** in the supplementary material. We plan to **release our maintained repository upon acceptance** to support future research and improvements. **References** [1] OverLoCK: An Overview-first-Look-Closely-next ConvNet with Context-Mixing Dynamic Kernels. Meng Lou, Yizhou Yu. [2] AnchorFormer: Differentiable Anchor Attention for Efficient Vision Transformer. Jiquan Shan, Junxiao Wang, Lifeng Zhao, Liang Cai, Hongyuan Zhang, Ioannis Liritzis. [3] BridgeDepth: Bridging Monocular and Stereo Reasoning with Latent Alignment. Tongfan Guan, Jiaxin Guo, Chen Wang, Yun-Hui Liu.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank the Area Chair and all reviewers for their careful evaluation, constructive feedback, and valuable suggestions.

We apologize for posting our responses relatively late in the discussion period. Over the past ten days, we have been training additional stereo-matching models with different monocular depth backbones and conducting further module analyses and comparative experiments to address the reviewers' concerns with concrete evidence. We have now provided detailed responses to each reviewer and incorporated the relevant experimental results, analyses, and methodological clarifications into the revised manuscript.

We greatly appreciate the reviewers' thoughtful comments, which have helped us improve the clarity, rigor, and experimental validation of ForceStereo. Although the discussion period is approaching its deadline, we remain fully committed to engaging actively with all reviewers. We warmly welcome any further questions, concerns, or requests for clarification, and we will make every effort to respond promptly before the discussion deadline.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.