acceptodds
Under review as a conference paper at ICLR 2027

MAGIC: Learning from Visibility Asymmetry for Unsupervised Stereo Matching

Abstract

Learning disparity in occluded regions remains difficult for unsupervised stereo matching. Photometric supervision lacks valid target-view correspondences in these regions, while the teacher and student in conventional binocular self-training share the same target view and therefore the same occlusions. Even when supervision is available, the small proportion of occluded pixels limits their contribution to training. We propose MAGIC, a multi-baseline geometric consistency framework for reliable and effective occlusion supervision. The teacher and student share a reference image but use different target views, allowing the teacher to observe correspondences that are occluded from the student. After aligning disparities across baselines, MAGIC uses predictions from teacher-visible regions to supervise student-occluded regions. An occlusion-aware weighting strategy strengthens supervision on teacher-visible but student-occluded pixels, preventing their training signal from being overwhelmed by non-occluded regions. We also introduce MBS20K, a synthetic multi-baseline stereo dataset spanning diverse scenes, weather, and lighting. Pre-trained on MBS20K, MAGIC generalizes to real-world datasets with consistently fewer occluded-region outliers. On KITTI, the pre-trained model already outperforms several fine-tuned unsupervised methods. Fine-tuning this model on standard binocular pairs achieves state-of-the-art unsupervised performance on KITTI 2015 and 2012. Incorporating synthesized multi-baseline views during fine-tuning further improves performance. Our code and dataset will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
stereo matching, disparity estimation

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

This paper introduces MAGIC, a multi-baseline framework for unsupervised stereo matching.

The main challenge for unsupervised stereo models consists of retrieving a proper supervisory signal in occluded regions, where the classical photometric consistency loss cannot help due to the lack of correspondences across views. To this end, MAGIC assumes the availability of multi-baseline data and exploits the different occlusion patterns occurring between the same reference image IrefI_{ref} and two distinct target views ItgtsI_{tgt}^s and ItgttI_{tgt}^t, collected respectively at baseline distances BsB^s and BtB^t.

MAGIC employs an EMA teacher-student scheme, in which the teacher and the students process the two stereo pairs Iref,ItgttI_{ref}, I_{tgt}^t and Iref,ItgtsI_{ref}, I_{tgt}^s, respectively. This allows the student to obtain reliable supervision on occluded regions in Iref,ItgtsI_{ref}, I_{tgt}^s that appear to be disoccluded in the Iref,ItgttI_{ref}, I_{tgt}^t pair, by properly identifying these regions and weighting the loss function there accordingly. To train stereo models with the MAGIC recipe, a new multi-baseline dataset rendered with CARLA is proposed (MBS20K).

Experimental results with different stereo backbones validate the effectiveness of the proposed approach under different settings: zero-shot generalization across datasets and weather conditions, and unsupervised fine-tuning on specific datasets. Some ablation studies are conducted to properly assess the impact of the different loss terms and design strategies.

Critical strengths

Two major strengths characterize this paper:

  • The main strength of this paper definitely lies in the quantitative results reported in Table 1. Indeed, state-of-the-art models like MonSter and FoundationStereo are slightly improved when trained in an unsupervised manner within the MAGIC framework, particularly on occluded regions. If the code and trained models are made publicly available upon acceptance, this would be the main contribution of this work.

  • A secondary strength of this work is the proposed MBS20K dataset: indeed, the effectiveness of MAGIC-trained models is probably also due to the quality of the data being rendered. In particular, the improvements across weather conditions shown in Tab. 2 probably stem from the diverse weather and lightning conditions exposed by MBS20K.

Critical weaknesses

The paper, unfortunately, exposes several critical weaknesses:

  • Although the proposed solution is sound and principled, it is not novel. The use of multi-baseline data for unsupervised or semi-supervised stereo has already been proposed by SMV (Yuan et al., 2021) and NeRFStereo (Tosi et al., 2023). Both introduce loss terms specifically designed to handle occlusions by exploiting photometric consistency across multiple target images compensating each others for occlusions.

  • A fair comparison against SMV and NeRFStereo should be provided to support that the proposed technique is more effective than both. Such a comparison is missing in the paper and is a critical omission: in Tab. 1, an entry for NeRFStereo is reported, using a different backbone (RAFT-Stereo) and without any SceneFlow initialization (which is, in turn, used by any MAGIC variant), while no entry for SMV is reported at all.

  • The need for the MBS20K dataset is unclear, and not sufficiently supported by experiments. NeRFStereo dataset already provides multi-baseline data with three different baselines, so using this dataset would have been a natural choice. SMV also provides multi-scopic data, still with constant baselines. A comparison between the same models trained on the different datasets is needed to validate the quality of the proposed MBS20K data.

  • Experiments reported in Tab. 3 are inherently flawed. Using MonSter weights to warp images and generate multi-baseline data breaks the unsupervised assumption, since MonSter is trained with ground-truth supervision. The whole comparison against unsupervised methods becomes, therefore, unfair.

Questions for authors

The critical weaknesses raised in the previous form must be addressed to make the paper acceptable at this venue.

Therefore, my main questions for the authors are:

  • Compare against NeRFStereo and SMV methods using the same backbone (MonSter or FoundationStereo), without any SceneFlow initialization, and on the same dataset, e.g., MBS20K. This is crucial to prove that the minimum photometric consistency loss proposed by the former, or the cross-view supervision proposed by the latter, is insufficient to handle occlusions during training.

  • Highlight the need for MBS20K by repeating the previous experiment with both NeRFStereo data and SMV data. This is crucial to prove that the MBS20K dataset plays a role in the results achieved, and to decouple the impact of the data from that of MAGIC.

  • For Tab. 3, either 1) just fine-tune the model on ground truth as usual and submit to the benchmark. This would prove that MAGIC allows for better pre-training, which transfers into better performance even after fine-tuning; or 2) generate the multi-baseline using MAGIC models themselves (or clarify if this is the case already). The simulated multi-baseline setup is flawed, since it stems from the availability of a model (MonSter) pre-trained with ground truth, making the whole process no longer unsupervised. Furthermore, DualNet (previous state-of-the-art in this experiment) also uses a different backbone (CFNet). A comparison using the same backbone would be appreciated

Post-discussion update

During the discussion period, the authors provided the requested missing comparisons with prior work, as well as addressed my concern about the unsupervised adaptation setting, proposing a new experiment replacing MonSter rendering with 4DGS-based rendering, which can be considered effectively unsupervised.

The authors also uploaded a revised version of the paper, implementing any of the changes that were discussed.

In light of this, I am glad to adjust my initial recommendation (Clear Reject) and increase it to Weak Accept.

8 comments

We thank the reviewer for recognizing the effectiveness of MAGIC, particularly its improvements in occluded regions, and the value of the MBS20K dataset. We also appreciate the constructive feedback, which helps improve the clarity and strengthen the experimental validation of our work. Below, we clarify the methodological contributions of MAGIC, provide controlled comparisons with existing multi-view methods, further examine the contribution of MBS20K, and address the concerns regarding target-domain fine-tuning.

Q1. Comparison with existing multi-view methods.

We acknowledge that SMV and NeRFStereo also exploit complementary visibility across multiple views to alleviate occlusion. Building on these works, the key distinction of MAGIC lies in how multi-baseline visibility complementarity is converted into progressively refined disparity supervision through an EMA teacher–student framework, explicit baseline-ratio alignment, and occlusion-aware weighting.

(1) Novelty clarification.

  • Visibility-asymmetric teacher–student learning. By sharing the same reference image but using different target views, the teacher and student naturally exhibit complementary visibility patterns. MAGIC exploits this visibility asymmetry: the teacher uses valid correspondence cues in its visible regions to provide reliable disparity supervision for student-occluded regions. As the student progressively improves its disparity completion capability in occluded regions, this capability is transferred back to the teacher through EMA updates, further improving the teacher's subsequent predictions.

  • Geometric disparity alignment. Both SMV and NeRFStereo adopt a fixed three-view configuration, with a central reference view and two target views on either side at equal baselines, ensuring that the resulting disparity maps share the same scale. In contrast, MAGIC imposes no such restriction on the relative positions of the reference and target views. By exploiting geometric consistency and rescaling teacher disparities according to the baseline ratio, MAGIC aligns disparity predictions from arbitrary reference–target pairs with different baselines. Consequently, an N-view set can generate up to N(N−1)2N(N-1)^2 teacher–student triplets, substantially expanding the available training samples.

  • Occlusion-aware weighting. As illustrated in Fig. 4, even with ground-truth supervision, the learning signal from occluded regions can be overwhelmed by the abundant co-visible pixels due to their highly imbalanced proportions. This imbalance limits the model's ability to learn accurate disparity completion in occluded regions. To address this issue, OAW explicitly increases the supervision weight of student-occluded but teacher-visible pixels, ensuring that their sparse but valuable supervision contributes more effectively to optimization.

(2) Quantitative comparison. We adopt CFNet as the shared backbone for MAGIC, SMV, and NeRFStereo. SMV predicts disparity and uncertainty through parallel heads attached to the generated cost volume, making CFNet's cost-volume filtering architecture a natural choice for preserving its original design. For a fair comparison of multi-view supervision, we use only NeRFStereo's triplet photometric loss, excluding its rendered disparity loss. All three methods use the same SceneFlow-pretrained CFNet initialization, learning rate, and iteration steps. Since the publicly provided SMV training data is no longer accessible, we evaluate all three methods on MBS20K and NeRFStereo datasets.

As shown in Tables I and II, MAGIC achieves the best overall performance among the three methods under the matched CFNet backbone and training configuration. On MBS20K, MAGIC achieves the best performance across all 16 reported metrics, outperforming both SMV and NeRFStereo. On the NeRFStereo dataset, MAGIC also outperforms SMV across all metrics and achieves better results than NeRFStereo on 14 out of 16 metrics. These results demonstrate the effectiveness of MAGIC under a unified backbone and initialization, providing a fair comparison with existing multi-view stereo methods.

Table I. Supplementary results to Table 1 in the main paper. Quantitative comparisons across different methods and datasets are provided.

MethodTraining DataKITTI 2015 OCCNOCALLKITTI 2012 OCCNOCALLMiddlebury OCCNOCALLETH3D OCCNOCALL
SMV-CFNetMBS20K20.239.389.6449.3811.6112.4953.6319.8924.3122.5413.8014.36
NeRFStereo-CFNetMBS20K11.786.106.2628.705.465.9944.5412.1316.5617.1112.0012.31
MAGIC-CFNetMBS20K10.954.294.4227.483.464.0035.5510.4713.978.662.933.23
SMV-CFNetNeRFStereo18.555.385.6832.755.055.6843.2013.1317.1735.9721.7922.46
NeRFStereo-CFNetNeRFStereo17.674.805.2029.263.984.5435.069.0512.6116.675.526.01
MAGIC-CFNetNeRFStereo13.184.684.8831.593.604.2428.468.1310.9313.465.335.74
MAGIC-IGEVMBS20K10.374.05‾\underline{4.05}4.21‾\underline{4.21}21.493.08‾\underline{3.08}3.48‾\underline{3.48}23.46‾\underline{23.46}6.34‾\underline{6.34}8.94‾\underline{8.94}7.982.51‾\underline{2.51}2.76‾\underline{2.76}
MAGIC-IGEVNeRFStereo10.93‾\underline{10.93}3.984.1521.56‾\underline{21.56}2.683.1122.145.487.848.17‾\underline{8.17}2.102.35

Table II. Supplementary results to Table 2 in the main paper. Quantitative comparisons across different methods and datasets are provided.

MethodTraining DataSunnyCloudyRainyFoggy
SMV-CFNetMBS20K8.198.2512.9112.29
NeRFStereo-CFNetMBS20K5.653.788.048.92
MAGIC-CFNetMBS20K1.971.604.682.61‾\underline{2.61}
SMV-CFNetNeRFStereo3.843.869.686.32
NeRFStereo-CFNetNeRFStereo3.083.686.794.23
MAGIC-CFNetNeRFStereo2.552.757.993.40
MAGIC-IGEVMBS20K2.15‾\underline{2.15}1.87‾\underline{1.87}5.07‾\underline{5.07}1.64
MAGIC-IGEVNeRFStereo2.702.397.522.61‾\underline{2.61}

Q2: Necessity of MBS20K.

As shown in Tables I and II, MAGIC trained on MBS20K outperforms its counterpart trained on the NeRFStereo dataset on most occlusion-region metrics and consistently achieves better results across all four weather conditions. We attribute these improvements to the complementary advantages of MBS20K. The NeRFStereo dataset relies on NeRF-based novel-view reconstruction, which may introduce reconstruction artifacts, particularly in occluded regions, potentially affecting the learning of occlusion completion. In contrast, MBS20K is collected using simulation engines, providing geometrically consistent multi-view observations while covering diverse weather conditions. These properties help improve stereo matching robustness in challenging occlusion regions and adverse weather. We further evaluate MAGIC with the IGEVStereo backbone and observe consistent trends.

Q3. Evaluation on the public KITTI benchmark.

(1) Clarification of the two fine-tuning settings. We first evaluate MAGIC under standard binocular unsupervised fine-tuning. This setting, using MAGIC-IGEV$^\ast$, provides a direct and fair comparison with existing unsupervised methods and achieves state-of-the-art performance on KITTI 2015 and 2012, demonstrating the effectiveness of our pre-training. We further explore whether additional multi-baseline observations can bring improvements during target-domain adaptation. Since KITTI does not provide multi-baseline images, we synthesize additional views through warping and inpainting using MonSter-estimated disparities. We acknowledge that MonSter introduces a supervised-model prior in constructing these input views. However, neither ground-truth disparities nor MonSter predictions are used as supervision targets during the subsequent MAGIC fine-tuning, which is optimized entirely with our unsupervised objectives. Thus, we consider MAGIC-IGEV$^\S$ an unsupervised fine-tuning setting, as the additional information is introduced only through the input views, without using ground-truth or MonSter-estimated disparities as supervision targets.

(2) Comparison with CFNet on KITTI 2015 benchmark. Following the reviewer's suggestion, we further fine-tune MAGIC-pretrained CFNet using ground-truth disparities from the KITTI 2015 training set. As shown in Table III, MAGIC-CFNet consistently outperforms the original CFNet across all six metrics. In particular, D1-all decreases from 1.88% to 1.76% in ALL regions and from 1.73% to 1.60% in NOC regions. These results demonstrate that the benefits of MAGIC pre-training persist after conventional supervised fine-tuning, providing a better initialization for subsequent target-domain adaptation.

Table III. Comparison on the KITTI 2015 benchmark.

MethodALL D1-bgALL D1-fgALL D1-allNOC D1-bgNOC D1-fgNOC D1-all
CFNet1.543.561.881.433.251.73
MAGIC-CFNet1.483.201.761.362.821.60
Reviewer_Cbi7OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies supervision for occluded regions in unsupervised stereo matching. MAGIC uses a teacher and a student that share a reference image but use different target views, allowing the teacher to observe regions occluded in the student view. Teacher disparities are then rescaled by the baseline ratio, and the loss weight for teacher-visible/student-occluded regions is increased. The paper also constructs MBS20K, a synthetic dataset with five-camera views, and evaluates zero-shot transfer on multiple real datasets and fine-tuning on KITTI.

Critical strengths
  1. Accurately identifies the lack of effective correspondence supervision in occluded regions for unsupervised stereo matching.
  2. Uses different target views to create visibility asymmetry, with a clear motivation.
  3. Provides clear geometric and optimization interpretations for baseline-ratio transformation and occlusion-aware weighting.
  4. MBS20K provides multi-baseline, multi-weather, and multi-scene training combinations.
Critical weaknesses
  1. Equation (4) is ambiguous: the text says that teacher-occluded pseudo-labels should be discarded, but the equation does not reflect this.
  2. The comparison with MBS20K ground-truth supervision covers only one OCC metric and does not sufficiently support a comprehensive comparison between unsupervised and supervised methods.
  3. The visibility masks contain false negatives and false positives, but the paper does not sufficiently analyze their effects on training and final performance.
  4. The gains from MBS20K data scale, weather diversity, and additional training are not fully disentangled from the core MAGIC mechanism.
Questions for authors
  1. Is the teacher-visibility factor missing from Equation (4)?
  2. Under the same backbone, MBS20K data, and other conditions, please compare ground-truth supervision with MAGIC unsupervision.
  3. How do precision/recall errors in the visibility masks affect the final results?
  4. Do the gains in generalization mainly come from the diversity of the proposed dataset or from the core mechanism of this paper?
5 comments

We thank the reviewer for recognizing our motivation, framework, and dataset. We also appreciate the constructive feedback, which helps improve the clarity and quality of our paper. Below, we address the concerns regarding Eq. (4), ground-truth supervision, visibility-mask reliability, and the respective contributions of MAGIC and MBS20K.

Q1. Is the teacher-visibility factor missing from Equation (4)?

We clarify that the teacher-visibility factor is not missing from Eq. (4). Specifically, the original weight map, A=Vt(Vs+ω(1−Vs))A=V^t(V^s+\omega(1-V^s)), can be expressed as:

\begin{cases} 0, & V^t = 0; \\ 1, & V^t = 1,\ V^s = 1,; \\ \omega, & V^t = 1,\ V^s = 0. \end{cases}$$ Thus, teacher-occluded regions are discarded, co-visible regions receive normal supervision, and student-occluded but teacher-visible regions receive increased supervision. **Q2. Comparison between ground-truth supervision and MAGIC.** Following the reviewer's suggestion, we provide a comprehensive comparison between **ground-truth-supervised IGEVStereo and MAGIC-IGEV, using the same backbone, SceneFlow initialization, and MBS20K training data.** As reported in Tables I and II, MAGIC-IGEV consistently outperforms its ground-truth-supervised counterpart across all metrics on KITTI 2015, KITTI 2012, Middlebury, and ETH3D, as well as all four weather conditions on DrivingStereo. Notably, MAGIC achieves consistent improvements in occluded regions, highlighting the effectiveness of our occlusion-aware weighting (OAW) strategy for disparity completion. **Table I. Supplementary results to Table 1 in the main paper.** | Method | Training data | KITTI 2015 | | | KITTI 2012 | | | Middlebury | | | ETH3D | | | | :------------------ | :------------ | ------------------: | -----------------: | -----------------: | ------------------: | -----------------: | -----------------: | ------------------: | -----------------: | -----------------: | -----------------: | -----------------: | -----------------: | | | | OCC | NOC | ALL | OCC | NOC | ALL | OCC | NOC | ALL | OCC | NOC | ALL | | IGEVStereo | SceneFlow | 14.26 | 5.60 | 5.79 | 33.66 | 4.92 | 5.59 | $\underline{24.28}$ | 7.25 | $\underline{9.91}$ | 9.76 | 4.06 | 4.39 | | IGEV+GT (Table 4.b) | MBS20K | $\underline{10.75}$ | $\underline{4.48}$ | $\underline{4.65}$ | $\underline{22.07}$ | $\underline{3.41}$ | $\underline{3.82}$ | 26.63 | $\underline{6.97}$ | 10.11 | $\underline{9.67}$ | $\underline{2.77}$ | $\underline{3.13}$ | | **MAGIC-IGEV** | MBS20K | **10.37** | **4.05** | **4.21** | **21.49** | **3.08** | **3.48** | **23.46** | **6.34** | **8.94** | **7.98** | **2.51** | **2.76** | **Table II. Supplementary results to Table 2 in the main paper.** | Method | Training data | Sunny | Cloudy | Rainy | Foggy | | :------------------ | :------------ | -----------------: | -----------------: | -----------------: | -----------------: | | IGEVStereo | SceneFlow | 4.59 | 5.15 | 15.47 | 4.49 | | IGEV+GT (Table 4.b) | MBS20K | $\underline{3.14}$ | $\underline{3.21}$ | $\underline{5.58}$ | $\underline{3.53}$ | | **MAGIC-IGEV** | MBS20K | **2.15** | **1.87** | **5.07** | **1.64** |

Q3. Effect of visibility-mask errors.

We acknowledge that the estimated visibility masks inevitably contain false positives (FPs) and false negatives (FNs). However, their effects are asymmetric. FPs incorrectly admit teacher-occluded pixels into geometric consistency supervision, potentially introducing erroneous gradients. In contrast, FNs discard valid teacher pseudo-labels, reducing supervision coverage. We consider FPs more critical, as incorrect supervision can be more harmful than missing supervision.

As reported in Appendix E (Table 8), our computed teacher visibility masks achieve 98.4%–98.8% precision across the six camera pairs, meaning that only 1.2%–1.6% of the admitted pixels are actually teacher-occluded. Our sensitivity analysis in Appendix F.1 further shows that a looser photometric threshold admits more teacher-occluded pixels (i.e., more FPs), introducing noisy supervision and degrading performance. We will further explore more accurate visibility mask estimation methods to reduce these errors.

Q4. Contribution of data diversity and the core MAGIC mechanism.

To disentangle the contributions of the proposed MAGIC mechanism and the training data, we evaluate five configurations, as reported in Tables III and IV: (a) the original IGEVStereo pretrained on SceneFlow; and MAGIC-IGEV trained on (b) full MBS20K (20,417 five-view sets), (c) the clear-noon subset (4,084 five-view sets), (d) 4,084 five-view sets randomly sampled from MBS20K, and (e) NeRFStereo dataset (26,528 seven-view sets synthesized from real-world monocular videos using NeRF reconstruction).

(1) Effectiveness and generalizability of MAGIC. To examine whether MAGIC's effectiveness depends on the scale and weather diversity of MBS20K, we first consider the reduced-data configurations (c) and (d). Despite using only a single weather condition or a smaller training set, both configurations outperform the original IGEVStereo (a) across all 16 metrics in Tables III and IV. Moreover, as shown in Tables I and II, MAGIC-IGEV trained on MBS20K also consistently outperforms its ground-truth-supervised counterpart. Together, these results demonstrate that MAGIC's improvements are not merely a consequence of the training data, but also stem from its multi-baseline geometric consistency framework. Furthermore, MAGIC-IGEV trained on NeRFStereo (e) also outperforms the original IGEVStereo (a), demonstrating that the effectiveness of MAGIC is not specific to MBS20K.

(2) Additional benefits of data diversity and scale. While MAGIC remains effective with reduced training data, greater weather diversity and data scale provide further benefits. Compared with (c), (d) reduces outliers from 6.01% to 5.38% on rainy scenes and from 1.88% to 1.67% on foggy scenes in Table IV, indicating improved generalization to adverse weather. Furthermore, compared with (d), (b) improves most metrics, demonstrating the additional benefit of larger-scale training data.

Table III. Supplementary results to Table 1 in the main paper. CN denotes the clear-noon subset of MBS20K. RS denotes a randomly sampled subset of MBS20K.

IDMethodTraining data (#sets × views)KITTI 2015KITTI 2012MiddleburyETH3D
OCCNOCALLOCCNOCALLOCCNOCALLOCCNOCALL
(a)IGEVStereoSceneFlow14.265.605.7933.664.925.5924.287.259.919.764.064.39
(b)MAGIC-IGEVMBS20K (20,417 × 5)10.374.05‾\underline{4.05}4.21‾\underline{4.21}21.493.08‾\underline{3.08}3.48‾\underline{3.48}23.466.34‾\underline{6.34}8.94‾\underline{8.94}7.982.51‾\underline{2.51}2.76‾\underline{2.76}
(c)MAGIC-IGEVMBS20K-CN (4,084 × 5)11.254.154.3223.693.123.5823.396.498.959.402.562.98
(d)MAGIC-IGEVMBS20K-RS (4,084 × 5)11.284.394.5822.323.163.5923.33‾\underline{23.33}6.869.078.542.48‾\underline{2.48}2.79
(e)MAGIC-IGEVNeRFStereo (26,528 × 7)10.93‾\underline{10.93}3.984.1521.56‾\underline{21.56}2.683.1122.145.487.848.17‾\underline{8.17}2.102.35

Table IV. Supplementary results to Table 2 in the main paper. CN denotes the clear-noon subset of MBS20K. RS denotes a randomly sampled subset of MBS20K.

IDMethodTraining data (#sets × views)SunnyCloudyRainyFoggy
(a)IGEVStereoSceneFlow4.595.1515.474.49
(b)MAGIC-IGEVMBS20K (20,417 × 5)2.151.87‾\underline{1.87}5.071.64
(c)MAGIC-IGEVMBS20K-CN (4,084 × 5)2.021.726.011.88
(d)MAGIC-IGEVMBS20K-RS (4,084 × 5)2.12‾\underline{2.12}1.965.38‾\underline{5.38}1.67‾\underline{1.67}
(e)MAGIC-IGEVNeRFStereo (26,528 × 7)2.702.397.522.61

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

The paper addresses disparity estimation in occluded regions for unsupervised stereo. The proposed method, MAGIC, provides the teacher and the student with the same reference image but different target views with different baselines, rescales the teacher disparity by the baseline ratio, and uses it as a pseudo-label. The authors also build MBS20K, a CARLA dataset recorded with a five-camera rig (20K frame sets, five weather/lighting conditions). Ablations and the appendix cover the loss terms, filtering and weighting, pseudo-label reliability, mask precision, hyperparameter sensitivity, and baseline-ratio perturbation.

Critical strengths

The motivation regarding failure in occluded regions for unsupervised stereo is spot on. The authors give the teacher a different target view to address the issue of the blind teacher. Quantitative results also show that binocular self-training on the same data shows worse results than the starting model, while MAGIC slightly outperforms training with ground truth on the same data in occluded regions. Appendix E also shows that the rescaled teacher labels are more accurate than the student in the regions the method targets.

Critical weaknesses

It is not very clear how much of the gain comes from the method and how much from the setup. For example, all models start from supervised SceneFlow weights, and Table 4 shows that training on MBS20K with ground truth already improves the depth of occluded regions on KITTI 2015. MBS20K is driving data that includes rain, so the large gains on KITTI and rainy DrivingStereo partly reflect domain similarity. Outside driving, on Middlebury and ETH3D, the gains are comparably small. In addition, the extra views for multi-baseline fine-tuning are generated from MonSter, a supervised stereo model, so those results are not purely unsupervised. Since the main claim is state-of-the-art unsupervised performance, these factors need to be separated more carefully.

Questions for authors
  1. Table 4, row (b), IGEVStereo trained on MBS20K with ground truth, is only reported on the KITTI 2015 occluded region. Please report the model trained on MBS20K with ground truth (Table 4, row b) in Tables 1 and 2, at least for IGEVStereo. At the moment, the zero-shot tables compare MAGIC only with models that never saw MBS20K, so it is hard to tell how much of the improvement comes from the supervision signal and how much from the data.

  2. Row (f) shows that training on the clear-noon subset still helps on KITTI 2015. Can the authors report the same clear-noon model on the rainy and foggy splits of DrivingStereo? This would show whether the large gains in rain in Table 2 come from the method or from the rainy scenes in MBS20K.

5 comments

We thank the reviewer for the positive assessment of our motivation, framework, and experimental validation. We also appreciate the reviewer’s constructive comments, which help improve the clarity and overall quality of our paper. Below, we address the concerns regarding the respective contributions of MAGIC and MBS20K, cross-weather generalization, and the role of MonSter-generated views in multi-baseline fine-tuning.

Q1. Contribution of the proposed method vs. the training data.

We agree that it is important to disentangle the contributions of the proposed method and the training data. We provide two complementary sets of experiments to address this concern.

(1) Controlled comparison on MBS20K. Following the reviewer's suggestion, we additionally evaluate IGEVStereo trained on MBS20K with ground-truth supervision. As shown in Tables I and II, MAGIC-IGEV consistently outperforms its gt-supervised counterpart. This extends the original comparison in Table 4(b) and demonstrates that the improvements cannot be attributed solely to the training data. These results highlight the effectiveness of MAGIC's multi-baseline geometric consistency framework, with OAW further improving disparity completion in occluded regions.

(2) Generalization beyond MBS20K and driving-scene similarity. To further examine whether MAGIC's effectiveness depends on MBS20K or its driving-scene distribution, we additionally train MAGIC-IGEV on the NeRFStereo dataset. Unlike MBS20K, NeRFStereo consists of real-world static indoor and outdoor scenes captured using a single handheld camera, from which seven-baseline views are synthesized through NeRF reconstruction. As shown in Tables I and II, MAGIC-IGEV achieves lower outlier rates than the original NeRFStereo model across all reported metrics, including substantial improvements on driving scenes. These results demonstrate the effectiveness of MAGIC across different scene distributions. We attribute this transferability to MAGIC's visibility-asymmetric EMA teacher–student framework and student-only appearance augmentation, which together encourage robust feature learning under challenging inputs.

Table I. Supplementary results to Table 1 in the main paper. Quantitative comparisons across different methods and datasets are provided.

MethodTraining dataKITTI 2015KITTI 2012MiddleburyETH3D
OCCNOCALLOCCNOCALLOCCNOCALLOCCNOCALL
IGEVStereoSceneFlow14.265.605.7933.664.925.5924.287.259.919.764.064.39
IGEV+GT (Table 4.b)MBS20K10.75‾\underline{10.75}4.484.6522.073.413.8226.636.9710.119.672.773.13
MAGIC-IGEVMBS20K10.374.05‾\underline{4.05}4.21‾\underline{4.21}21.493.08‾\underline{3.08}3.48‾\underline{3.48}23.46‾\underline{23.46}6.34‾\underline{6.34}8.94‾\underline{8.94}7.982.51‾\underline{2.51}2.76‾\underline{2.76}
NeRFStereoNeRFStereo14.625.235.4326.973.514.0431.106.7910.388.352.783.07
MAGIC-IGEVNeRFStereo10.933.984.1521.56‾\underline{21.56}2.683.1122.145.487.848.17‾\underline{8.17}2.102.35

Table II. Supplementary results to Table 2 in the main paper. Quantitative comparisons across different methods and datasets are provided.

MethodTraining dataSunnyCloudyRainyFoggy
IGEVStereoSceneFlow4.595.1515.474.49
IGEV+GT (Table 4.b)MBS20K3.143.215.58‾\underline{5.58}3.53
MAGIC-IGEVMBS20K2.151.875.071.64
NeRFStereoNeRFStereo2.882.9110.203.93
MAGIC-IGEVNeRFStereo2.70‾\underline{2.70}2.39‾\underline{2.39}7.522.61‾\underline{2.61}

Q2. Generalization to unseen weather conditions.

To investigate whether the improvements on rainy and foggy scenes rely on exposure to similar weather conditions during training, we evaluate MAGIC-IGEV trained exclusively on the clear-noon subset of MBS20K.

As shown in Table III, MAGIC-IGEV-CN consistently outperforms the original IGEVStereo. Notably, despite being trained only on clear-noon scenes, it reduces the Out-3 error by 61.1% on rainy scenes (15.47% → 6.01%) and 58.1% on foggy scenes (4.49% → 1.88%). These results demonstrate that MAGIC generalizes effectively to unseen weather conditions, without relying on weather-specific training samples. Furthermore, the full MAGIC-IGEV model achieves lower errors on rainy and foggy scenes, showing that incorporating weather-diverse data in MBS20K provides additional benefits.

Table III. Supplementary results to Table 2 in the main paper. CN denotes the clear-noon subset of MBS20K.

MethodTraining dataSunnyCloudyRainyFoggy
IGEVStereoSceneFlow4.595.1515.474.49
MAGIC-IGEVMBS20K2.15‾\underline{2.15}1.87‾\underline{1.87}5.071.64
MAGIC-IGEV-CN (Table 4.f)MBS20K-CN2.021.726.01‾\underline{6.01}1.88‾\underline{1.88}

Q3. Clarification on multi-baseline fine-tuning and the unsupervised SOTA claim.

We apologize that the distinction between the two fine-tuning settings was not sufficiently clear in the original manuscript. Our unsupervised SOTA claim is based on MAGIC-IGEV$^\ast$, which uses standard binocular KITTI pairs for fine-tuning (Section 4.3, Line 426).

MAGIC-IGEV$^\S$ is an additional experiment to investigate the potential benefit of multi-baseline observations for target-domain adaptation. We acknowledge that this setting uses additional input views compared with standard binocular methods. Since KITTI does not provide multi-baseline images, we use MonSter solely to help synthesize additional views. Importantly, the subsequent MAGIC fine-tuning remains unsupervised, without using ground-truth disparities or MonSter predictions as output supervision.

Moreover, the synthesized views are imperfect due to errors in disparity estimation, warping, and inpainting. Nevertheless, MAGIC-IGEV$^\S$ achieves further improvements over standard binocular fine-tuning. This suggests that real-world multi-baseline observations could provide additional benefits for target-domain adaptation.

Update to Q3.

We appreciate the reviewer’s earlier concern regarding the use of MonSter-generated views. Upon further consideration, we recognize that although MonSter predictions are not used as explicit supervision targets, they determine the geometry of the synthesized input views and may therefore introduce some supervised stereo prior into the training process. To avoid this concern, we will remove the original MAGIC-IGEV$^\S$ results from Table 3 and the corresponding view synthesis pipeline in Appendix C.

To further investigate the benefits of multi-baseline observations for unsupervised target-domain adaptation without relying on a supervised stereo model, we redesign this experiment using 4D Gaussian Splatting (4DGS), without ground-truth point-cloud initialization or depth/disparity supervision. Specifically, we divide KITTI Odometry sequences into multiple clips and independently reconstruct each clip using 4DGS. Each clip contains 50 synchronized stereo pairs, of which 40 are used for reconstruction and the remaining 10 are excluded from reconstruction and used to evaluate reconstruction quality. We then render four additional views at different camera positions for each timestamp, forming six-view multi-baseline sets. After temporal subsampling and quality filtering, we obtain 1,912 multi-baseline sets for fine-tuning MAGIC-IGEV pre-trained on MBS20K.

As shown in Table IV, the resulting MAGIC-IGEV$^\diamond$ consistently outperforms both the standard binocular fine-tuning variant MAGIC-IGEV$^\ast$ and the previous state-of-the-art method DualNet across all metrics. We hope this new experiment can replace the original MonSter-based view synthesis experiment and provide further evidence of the effectiveness of MAGIC for target-domain adaptation. We will also publicly release the constructed multi-baseline dataset to facilitate research in the community.

Table IV. Supplementary results to Table 3 in the main paper. ∗^* denotes binocular fine-tuning, ⋄^\diamond denotes multi-baseline fine-tuning.

MethodKITTI 2015 (D1)KITTI 2012 (Out-3)
NOCALLNOCALL
DualNet2.672.922.062.59
MAGIC-IGEV$^*$2.592.792.022.36
MAGIC-IGEV$^\diamond$2.352.511.842.21

We thank the reviewer again for the valuable comments and will submit the revised manuscript for further review as soon as possible.

Comments on the paper

2 comments

We sincerely thank the Area Chair and all reviewers for their time, patience, careful evaluation, and constructive suggestions. Their feedback has helped us clarify the methodological contributions of MAGIC, better disentangle the effects of the proposed framework and training data, and strengthen our experimental validation. We have incorporated the relevant additional experiments, analyses, and clarifications into the revised manuscript.

We are grateful for the reviewers' thoughtful engagement and hope that the revisions adequately address their concerns. We welcome any further questions or requests for clarification during the discussion.

Dear authors,

I reviewed the revised version of the paper, and I believe it successfully addresses the points that were discussed in our previous iterations. I also checked the other reviews and the changes to the manuscript made to address other concerns, and I could not find any other critical factor that could be grounds for rejecting the paper.

Therefore, I confirm my intention to increase my final score to Weak Accept.

Kind regards,

Reviewer xPjc

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.