acceptodds
Under review as a conference paper at ICLR 2027

NBS: No Bias Stereo

Abstract

Stereo reconstruction is one of the last remaining computer vision tasks where state-of-the-art methods still rely on heavy, stereo-specific architectural inductive biases. Although it has been demonstrated that the task can be solved using general-purpose methods, such methods have since fallen behind specialized models, and it is widely believed that inductive biases in stereo are strictly necessary for both high-quality results and computational efficiency. We challenge this paradigm. In this paper, we demonstrate that both state-of-the-art accuracy and superior runtime efficiency are achievable with a model devoid of stereo-specific architectural inductive biases, relying instead on a simple, end-to-end Vision Transformer. By training on large-scale synthetic corpora, we show that a data-driven approach can surpass explicitly engineered geometry: our model ranks first on the ETH3D benchmark, leads on most metrics across four other benchmarks, and runs 4 faster than the previous best method. These results suggest that stereo-specific inductive biases are no longer a prerequisite for stereo matching, paving the way for scalable, general-purpose models for stereo reconstruction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Stereo, 3D Vision, Vision Transformer

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dkDXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents NBS, a stereo matching model built on a DINOv2-initialized ViT-Large with alternating within-view and cross-view attention and a DPT disparity prediction head. The method combines high-resolution learnable positional embeddings, standard disparity and gradient-matching losses, and large-scale synthetic training. It reports strong accuracy on ETH3D, Middlebury, and two additional evaluation datasets, together with substantial inference speedups over selected stereo baselines. The main claim is that explicit stereo-specific matching and refinement modules are unnecessary for achieving state-of-the-art performance.

Critical strengths
  1. NBS achieves excellent results on ETH3D, including a substantial reduction in strict disparity errors relative to recent methods. The results demonstrate that the proposed pipeline can achieve high reconstruction accuracy without explicit cost volumes or recurrent refinement. The reported combination of accuracy and efficiency is practically relevant.

  2. NBS-Base, trained exclusively on public stereo data, achieves the best EPE on SimpleProc-S (0.25 versus 0.29 for S2M2). This is useful evidence that the approach's performance cannot be attributed solely to the additional proprietary high-resolution training data.

Critical weaknesses
  1. The central conceptual claim is substantially anticipated by prior work, and the incremental scientific contribution remains insufficiently established. CroCo v2 already achieved state-of-the-art stereo performance without cost volumes or iterative refinement, while DA3/VGGT established the single-pretrained-ViT paradigm with within-view/cross-view attention closely followed by NBS. The key question is therefore what NBS establishes beyond improving the performance of this existing paradigm. While the evaluation is extensive, they not entirely resolve this. In addition, comparing stereo-pair counts does not disentangle architecture from training exposure, initialization, optimization, or resolution, despite the claim in Appendix A.5. The positional-embedding ablation establishes an improvement on FSD, but does not establish its contribution to the main cross-domain results. I acknowledge the strong system performance; however, the manuscript does not sufficiently establish a distinct, transferable finding.

  2. The main ETH3D and Middlebury results rely on benchmark-specific fine-tuning, which does not directly establish zero-shot stereo generalization. Although Figure 1 provides a qualitative zero-shot example, quantitative zero-shot disparity evaluations on these standard real-world benchmarks are not clearly reported. The high-resolution results are also obtained using separately adapted models with an expanded training mixture, making it difficult to distinguish the effects of increased resolution, additional training data, and the proposed design choices. Furthermore, the reported fourfold inference speedup is measured at 966×546 resolution, while the corresponding accuracy-latency trade-off at higher resolutions is not established. The SimpleProc and XYZ-IBD evaluations provide useful additional evidence, but the former remains synthetic and the latter contains only 55 selected stereo pairs, limiting conclusions about broader real-world transfer. They leave uncertainty about how broadly the advantages extend beyond the particular evaluation settings. Given the limited methodological novelty, I consider a clearer characterization of these empirical advantages important for assessing the paper's contribution.

Questions for authors
  1. What is the main contribution or findings beyond CroCo v2, VGGT, and DA3, particularly given the similar attention architecture used in DA3?
  2. Can the authors report zero-shot NBS-Base results on ETH3D and Middlebury and clarify the fine-tuning conditions of the baselines?
  3. Can the authors provide training exposure, compute, and dataset sampling ratios? What evidence supports the broader architecture-versus-scaling conclusion?
  4. Does the high-resolution PE improve performance on independent real-world data? How do NBS-HighRes latency and memory compare with baselines, including Fast-FoundationStereo?
  5. How are depth scale and camera calibration handled for DA3/VGGT in Table 4, particularly given the unusually large AbsRel values for DA3Large?
  6. How were the 55 XYZ-IBD pairs selected, and how similar are these scenes to the proprietary training data?
4 comments

Thank you for the thorough review. The shared Table G1 and Table 5 are in the general response and the revised main paper, respectively; the public-only high-resolution result is in our reply to Reviewer EHsk (W2/Q1).

Q1, contribution. We show that stereo-specific modules are not a prerequisite for top-tier stereo matching: a single strong counter-example suffices—a standard ViT+DPT with no cost volume, 3D convolution, or iterative refinement can lead in both accuracy and speed on ETH3D and at 966×546966\times 546. Specifically, it:

  • ranks #1 on ETH3D, trained with public data only;
  • achieves lower zero-shot EPE and [email protected] than S2M2 and FS on ETH3D, even though both baselines were trained on it;
  • is both the fastest and the most accurate method at 966×546966\times 546 (Q4b).

Apart from ours, every non-anonymous top-30 method on ETH3D and Middlebury relies on stereo-specific modules. CroCo-Stereo showed that cost-volume-free stereo is feasible, but it trails specialized methods substantially (ETH3D bad@1: 1.14 vs. 0.16). Without rectified disparity formulation, DA3 and VGGT do not reach stereo-level metric accuracy (Table 4; scale-aligned geometry is compared in Table 3). Because the architecture is generic, NBS directly leverages standard ViT optimizations such as FlashAttention and F16 precision, driving its runtime advantage.

Q2, zero-shot and baseline conditions. Table G1 in the general response reports quantitative zero-shot evaluations: neither NBS-Base nor NBS-HighRes ever saw ETH3D or Middlebury during training, whereas S2M2 and FS were trained on both training splits (per their authors). Even so, zero-shot NBS-Base outperforms both on ETH3D across all four metrics (EPE 0.138 vs. 0.190 and 0.171). Regarding our variants: NBS-ETH3D fine-tunes NBS-Base on FSD + ETH3D + Middlebury; NBS-HighRes adapts NBS-Base to higher resolution on the 14-dataset mixture in Section 3.4.2; and NBS-Middlebury fine-tunes NBS-HighRes with Middlebury added.

Q3, exposure and controlled factor ablations.

  • Base stage: 170k steps at batch size 640 (128 A100s, bf16) on FSD (1.11M pairs).
  • High-res stage: initialized from the base stage on 14 datasets (3.47M public + 2.40M internal pairs), followed by a final fp32 stage on 64 H100s.
  • Controlled ablations: Table 5 (Section 4.5) isolates initialization, pretraining type, optimizer, data fraction, and sample budget. Because our claim is that a generic ViT+DPT architecture is sufficient, we hold the architecture fixed rather than ablating architectural variants.

Q4a, does high-resolution PE help on real data? Yes—we added a zero-shot real-world PE ablation on ETH3D and Middlebury in Appendix A.5 (Table 10). The results confirm that high-resolution PEs consistently lower EPE and [email protected] across scales on real-world scenes.

Q4b, latency and memory. Additional efficiency analysis is provided in the revised Section 4.1.3, Table 2, and Figure 4, with 1932×10921932\times 1092 results in Appendix A.7 (Table 12). At 966×546966\times 546, NBS is both the fastest and the most accurate method. At 1932×10921932\times 1092, NBS matches the latency of lightweight Fast-FoundationStereo (0.51 s vs. 0.50 s) with 3.8×3.8\times lower peak memory (6.56 GB vs. 24.60 GB), while running 1.6×1.6\times and 4.7×4.7\times faster than S2M2 and FS and achieving the best accuracy.

Q5, Table 4. We thank the reviewer for catching this. DA3 and VGGT predict depth only up to an unknown global scale. For Table 4, we recover metric scale from the predicted two-view relative pose: we compute the predicted camera baseline ∥C0−C1∥\|C_0 - C_1\| and rescale depth by Bgt/∥C0−C1∥B_{\text{gt}} / \|C_0 - C_1\| using the known rig baseline BgtB_{\text{gt}}. Neither DA3-Large nor VGGT is given GT intrinsics: VGGT does not accept them, and DA3 uses intrinsics only together with input extrinsics (pose-conditioned mode), which we avoided because DA3's built-in Umeyama scale alignment is degenerate for two views (DA3 GitHub issue #241). DA3Metric, a separate metric-depth model, is converted via depth=f⋅d^/300\text{depth} = f \cdot \hat{d} / 300 using the GT focal length ff, as specified by its authors. While verifying this, we found a unit-conversion bug in our DA3-Large script (a spurious ×1000\times 1000 applied after baseline rescaling), which caused the anomalous AbsRel values. The corrected AbsRel values are 0.886 / 2.609 / 1.342 (ETH3D / SimpleProc-S / XYZ-IBD), now updated in Table 4. The VGGT entries are correct; its SimpleProc-S AbsRel of 79.9 stems from near-zero predicted baselines. The remaining gap reflects the ill-conditioned scale estimate when inferring a narrow, nearly pure-translation stereo baseline from two close views, motivating direct disparity prediction under known rectification as in NBS.

Q6, XYZ-IBD.

  • Selection: The released raw data contains 60 scenes; 55 have valid ground-truth depth, and the other five were excluded due to ground-truth artifacts. In each scene, we pick the two frames closest to a rectified stereo pair that both contain the full bin, and fuse five structured-light scans into the ground truth.
  • Overlap with training data: Part of our internal synthetic dataset contains top-down views and industrial parts, but none of the XYZ-IBD objects, and the XYZ-IBD camera intrinsics and viewpoints were not replicated.

Thank you for the detailed response and additional experiments. The added results address several of my concerns by strengthening the evidence for NBS’s accuracy, efficiency, and zero-shot generalization. I also appreciate the real-world PE ablations and the clearer characterization of the system’s practical advantages.

My remaining reservation concerns the significance of the advance over prior work. The architecture, losses, and acceleration techniques largely build on established methods. The PE-related findings provide useful design insights, but I view the main contribution as a strong empirical realization of established design principles, with a more limited methodological or explanatory advance. Limited methodological novelty does not by itself preclude acceptance, and I consider the strong empirical results an important contribution, so I would like to decide my final evaluation after dicussion with other reviewers.

Thank you for the thoughtful follow-up and constructive feedback.

We intentionally kept our model simple and general. By avoiding new custom modules, we show that high-quality stereo reconstruction does not require stereo-specific components.

Reviewer_uGa7OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is about stereo vision reconstruction with a neural network. The authors propose to use a large-scale Transformer-based stereo architecture aiming to avoid explicit cost volumes or iterative disparity refinement. Building on DINOv2 with registers, a standard Vision Transformer backbone (ViT-Large) is used leveraging the original pretrained weights from DINOv2. The encore is followed by a DPT disparity prediction decoder.

The approach has been validated on the standard Middlebury and ETH3D benchmarks and compare favourably with prior methods.

Critical strengths

The paper show the effectiveness of a particular large-scale Transformer-based stereo architecture.

The experimental results are convincing and the performance improvements are meaningful, particularly on ETH3D and the synthetic evaluation datasets.

Critical weaknesses

The idea of combining Transformer-based feature processing and a DPT head to regress disparities was already proposed in CroCo v2. The proposed approach has some architectural differences but is difficult to understand if the performance improvements come from the architecture, DINOv2 pretraining, the amount of training data, or the training strategy. Indeed, without ablation studies the performance gains cannot be attributed to a particular contribution.

The proposed method predicts dense disparities even in regions where stereo correspondence is fundamentally ambiguous (e.g. textureless or occluded surfaces). Although the authors present this capability as an advantage, they do not investigate whether the resulting predictions correspond to observable geometry or represent hallucinations induced by learned monocular priors.

Questions for authors

Can the authors should clarify the architectural novelty relative to CroCo v2 ?

Can the authors and provide controlled ablation studies to distinguish the benefits of the architecture from those of pretraining, training-data scale, and optimization ?

Can the author provide a dedicated evaluation of textureless regions, geometric consistency, or predictive uncertainty ?

3 comments

Thank you for the positive assessment. The shared Table G1 and Table 5 are in the general response and the revised main paper, respectively.

Q1, CroCo v2. The key difference is generality. Architecturally, CroCo's decoder is built for two views in fixed asymmetric roles: one view queries the other through cross-attention, and only that view is updated. NBS doesn't have any stereo-specific architectural inductive biases. NBS applies standard global self-attention across the tokens of all input views symmetrically, extending naturally to more views without architectural changes, as in VGGT and DA3. For initialization, NBS requires no stereo/cross-view pretraining (CroCo relies on cross-view completion). At inference, NBS runs in a single feed-forward pass rather than tiling 704×352704\times 352 crops, running 12×12\times faster at 966×546966\times 546 (Table 2).

W1 / Q2 (controlled ablations). Table 5 in the revised main paper isolates pretraining (row 2), pretraining type (row 3), optimizer (row 4), data volume (row 5), and training budget (row 6). Task-agnostic initialization, training budget, and the optimizer each contribute substantially, whereas geometry-specialized pretraining (DA3-Large) gives no advantage. Because our claim is that an off-the-shelf ViT+DPT architecture suffices without stereo-specific modules—rather than proposing a new ViT variant—we hold the architecture fixed in this ablation.

W2 / Q3, ambiguous regions. We examine this from four angles:

  • Use of the second view: On ETH3D, NBS achieves AbsRel 0.021 with stereo input versus 0.164 with the right image blanked (even after least-squares scale-and-shift alignment; Tables 3–4). This 8×8\times gap confirms that NBS actively relies on stereo correspondence rather than monocular priors when both views are present.
  • Occlusions: Moving from non-occluded to all pixels on the ETH3D test set, bad@1 increases from 0.14→0.160.14 \rightarrow 0.16 for NBS, compared to 0.22→0.260.22 \rightarrow 0.26 for S2M2 and 0.26→0.480.26 \rightarrow 0.48 for FS (Table 8). NBS thus degrades the least in occluded regions lacking correspondence.
  • Low-gradient regions: Following Scharstein & Szeliski (2002), we mark non-occluded pixels as textureless when the mean squared horizontal gradient (I(x,y)−I(x−1,y))2(I(x,y) - I(x-1,y))^2 on [0,255][0, 255] grayscale is below 4.04.0 over an 11×1111\times 11 window (ETH3D) or 21×2121\times 21 window (Middlebury). Because this mask selects smooth interior surfaces whereas high-gradient pixels include sharp depth discontinuities (where large boundary errors concentrate), EPE on Middlebury is naturally lower on low-gradient pixels across all methods:
EPE / [email protected] / bad@2ETH3D low-grad. (35%)ETH3D high-grad.Midd-F low-grad. (51%)Midd-F high-grad.
NBS-HighRes0.129 / 2.22 / 0.010.143 / 1.87 / 0.520.648 / 32.70 / 4.931.252 / 37.05 / 5.36
S2M20.200 / 4.22 / 0.230.169 / 2.27 / 0.400.425 / 23.15 / 1.230.838 / 21.56 / 2.74
FS0.155 / 4.35 / 0.110.143 / 1.95 / 0.350.403 / 19.60 / 1.090.993 / 20.78 / 3.59

On ETH3D, S2M2 and FS have higher EPE on low-gradient pixels than on high-gradient pixels, whereas NBS achieves lower EPE on low-gradient pixels (0.129 vs. 0.143), showing no sign of prior-induced hallucination errors. On Middlebury, NBS trails overall, but the ratio of low-gradient to high-gradient EPE (≈0.5×\approx 0.5\times) is nearly identical across all three methods (even though S2M2 and FS were trained on Middlebury), indicating dataset/resolution difficulty rather than monocular prior error.

  • Uncertainty: NBS predicts point disparities without an uncertainty output.
Reviewer_uGa7OpenReview ↗

Answer to the answer

The authors have answered some of my concerns and confirmed my initial positive evaluation.

Thank you for engaging with our response and confirming your positive assessment!

Please let us know if any further concern remains.

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

The paper challenges the prevailing view that stereo matching requires specialized architectural inductive biases (cost volumes, 3D convolution, iterative GRU refinement). It presents NBS, a plain end-to-end Vision Transformer (ViT-Large initialized from DINOv2 with registers, local attention in layers 1–7 and global attention in alternating layers 8–24, a high-resolution positional-embedding grid of 148×148 / 296×296, and a DPT decoder) that contains no explicit cost volume, 3D conv, or recurrent refinement. Trained on large-scale synthetic corpora (a base stage on FSD alone at 966×546 for 170k steps on 128×A100, then a high-resolution adaptation on a 2.4M-scene proprietary synthetic set plus 13 public datasets at 1932×1330 on H100s), the model ranks 1st on ETH3D, reaches SoTA on 4 of 5 evaluated benchmarks (ETH3D, SimpleProc-S/M, XYZ-IBD) and is competitive on Middlebury, while running ~4× faster than the prior best (S2M2) and ~14.5× faster than FoundationStereo in F16. The authors also show an emergent monocular-stereo fusion: with the right image blanked, NBS still produces plausible depth competitive with dedicated monodepth models (DA3, VGGT).

Critical strengths
  1. Impressive, practically relevant empirical result. NBS achieves top-tier accuracy on ETH3D and most other benchmarks while being substantially faster (4× vs S2M2, 14.5× vs FoundationStereo in F16 at matched 966×546). This is a convincing demonstration that the "bitter lesson" applies to stereo and a meaningful step toward general-purpose reconstruction models.
  2. Clean, reproducible-as-far-as-possible protocol for the public part. The base model and ETH3D submission use only public data; the paper provides SimpleProc generation seeds and the post-processed XYZ-IBD set, plus a thoughtful positional-embedding resolution ablation (148×148 cuts [email protected] by 7.21 pts vs 37×37).
Critical weaknesses
  1. Limited and partly overstated methodological novelty. The architecture is essentially off-the-shelf DINOv2 + DPT decoder — the same minimalist recipe already used by DA3, VGGT, and MoGe for geometry. The real advances are scale, high-res PE, and tuning, not a conceptual break. The "devoid of inductive biases" framing is also imprecise: the model inherits massive pretraining bias from DINOv2; the claim is only "no stereo-specific architectural bias." A direct comparison to a simply fine-tuned DA3/VGGT stereo baseline is missing and would clarify the marginal contribution.
  2. Reproducibility and fairness hinging on undisclosed data. Because the strongest SoTA claims (Middlebury, SimpleProc-M, XYZ-IBD) depend on the unreleased 2.4M internal synthetic set, the community cannot reproduce them, and fair benchmarking against specialized models trained on public data is complicated. Combined with the Middlebury downsampling caveat and the tiny XYZ-IBD set, the breadth of the "SoTA" claim is weaker than the abstract suggests.
Questions for authors
  1. How much of the gains on Middlebury / SimpleProc-M / XYZ-IBD come from the unreleased 2.4M internal dataset? Please report results for the publicly-trained base model on these benchmarks to quantify the internal-data contribution.
  2. For Middlebury, can you estimate how much EPE/RMS the 1932×1330→upsample bottleneck costs, or restrict the comparison to methods also evaluated at that resolution? As written, the Middlebury number is not directly comparable to native-resolution submissions.
  3. Beyond "no stereo-specific bias," how does NBS compare to fine-tuning an existing vanilla-DINO geometry foundation model (DA3/VGGT) directly for stereo? This would isolate whether the contribution is the architecture or the large-scale stereo training.
  4. XYZ-IBD has only 55 pairs with structured-light-derived GT (median fusion of 5 scans, 2 selected views). Given this small, reconstructed ground truth, how stable are the conclusions, and is any cross-validation performed?
1 comment

Thank you for the careful review. The shared Table G1 and Table 5 are in the general response and the revised main paper, respectively.

W1 (wording and novelty) We agree on the wording: our claim is "no stereo-specific architectural inductive bias," and DINOv2 pretraining is part of the recipe (in Table 5, random initialization gives 1.6×1.6\times higher EPE). We use this precise wording throughout the revision. Regarding novelty, the architecture is intentionally standard, following the layout of VGGT / DA3. Our contribution is demonstrating that this generic recipe suffices to rank #1 on ETH3D using public data alone, which no prior method without stereo-specific modules had achieved.

Q3 (fine-tuning DA3 / VGGT for stereo). We ran this with DA3-Large, fine-tuned for stereo under our recipe (Table 5, row 3). Its EPE is 18% higher than with DINOv2 initialization across all three datasets, showing that starting from a geometry foundation model gives no advantage here.

W2 / Q1 (internal data). With public data only, NBS ranks first on the ETH3D test benchmark (Table 1), and zero-shot NBS-Base leads S2M2 and FS on the ETH3D training split even though both were trained on it (Table G1). To measure the influence of internal data at high resolution, we fine-tuned NBS-Base at 1932×10921932\times 1092 using only the public datasets listed in Section 3.4.2 (reporting EPE / bad@1):

ModelSimpleProc-MXYZ-IBDMidd-F
NBS, public only0.439 / 6.119.73 / 58.31.229 / 20.6
NBS-HighRes (public + internal, full schedule)0.40 / 4.2211.39 / 55.61.158 / 15.5
S2M20.444 / 6.0519.24 / 57.20.831 / 8.70
FS0.612 / 6.2715.75 / 54.71.015 / 9.45

The public-only model matches S2M2 on SimpleProc-M and achieves the lowest EPE on XYZ-IBD, showing that competitive high-resolution accuracy does not require proprietary data. It does not reach our full NBS-HighRes margin on SimpleProc-M and Middlebury; that gap reflects both the internal data and NBS-HighRes's longer, progressive training schedule. We clearly mark which evaluations use internal data in Section 3.3 and Table 11.

W2 (reproducibility). While we cannot release the code, weights, or proprietary dataset, the public-only results can be reproduced from public data alone: they require only FSD and the official benchmark training sets, with the full architecture and training recipe documented in Section 3. We also provide the SimpleProc generation seeds and the post-processed XYZ-IBD evaluation set in the supplement.

Q2, Middlebury resolution.

  • Our Middlebury entry is predicted at 1932×13301932\times 1330 and upsampled (explicitly noted in Figure 1, Section 4.1, and Table 1), so it is not directly comparable with native-resolution entries; this bottleneck works against NBS.
  • Resolution matters on high-resolution inputs: while no NBS model runs at native 3000×20003000\times 2000 scale, simply downsampling the Middlebury ground truth to 1932×13301932\times 1330 and upsampling it back (with no stereo matching involved) already incurs an EPE of 0.157 px, and S2M2 loses 14% EPE (0.558→0.6370.558 \rightarrow 0.637) under the same down- and upsampling.
  • For context, the leaderboard gap between our model and S2M2 is 0.05 px (EPE 1.05 vs. 1.00), and NBS achieves the best RMS (7.21), though we do not claim that NBS would lead at native resolution.

Q4, XYZ-IBD.

  • Selection: The released raw dataset has 60 scenes; 55 have valid ground-truth depth, and the remaining five were excluded due to ground-truth artifacts. In each scene, we select the two frames closest to a rectified stereo pair that both contain the full bin, and fuse five structured-light scans into the ground truth. The resulting 55 pairs in XYZ-IBD are 1.57×1.57\times larger than the Middlebury and ETH3D test sets combined (15 and 20 pairs, respectively).
  • Cross-validation: Because no model is trained or tuned on XYZ-IBD (strictly zero-shot evaluation), cross-validation splits are not needed.
  • Results: NBS achieves the lowest EPE, bad@2, and bad@4, while FS leads on [email protected]/bad@1 and CREStereo on RMS (Table 6). We use XYZ-IBD as real-world evidence complementary to ETH3D.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their careful and constructive reviews. We ran new evaluations and controlled training ablations, and have uploaded a revised PDF.

Clarifications in the revision

  1. We consistently state "no stereo-specific architectural inductive bias" and name DINOv2 pretraining upfront as part of the recipe.
  2. We add Fast-FoundationStereo to our efficiency evaluation in Section 4.1.3, Table 2, and Figure 4, and add high-resolution (1932×10921932\times 1092) efficiency analysis in Appendix A.7 (Table 12).
  3. We corrected the DA3-Large values in Table 4.
  4. We add a controlled ablation study with down-scaled training in Section 4.5 (Table 5).
  5. We add a real-world ablation study on high-resolution positional embeddings in Appendix A.5 (Table 10).

Protocol for the new numbers

We use a single evaluation harness, which reproduces the paper's numbers. ETH3D and Middlebury use their training splits (27 and 15 pairs) at full resolution, scored on all pixels.

Table G1: Zero-shot evaluation of our NBS models on the ETH3D and Middlebury training splits.

ModelETH3D EPE[email protected]bad@1bad@2Midd-F EPEMidd-F bad@2
NBS-Base (FSD only)0.1382.280.610.351.65212.52
NBS-HighRes0.1391.970.670.391.1587.19
S2M2 (trained on ETH3D/Midd.)0.1903.380.800.480.8313.64
FS* (trained on ETH3D/Midd.)0.1712.820.670.441.0154.37
*Evaluated at 1932×13301932\times 1330: FS runs out of memory at native Middlebury resolution on a 40 GB A100.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.