{"id":"9ea454d0-561d-4561-8851-d83fd8b9baea","arxiv_id":"2506.04908","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using 3D Gaussian Splatting renderings with FoundationStereo pseudo-depth labels produces synthetic stereo training data that transfers to real-world zero-shot benchmarks as well as or better than prior NeRF-based methods.","lead":"The authors generate fake stereo image pairs by rendering 3D Gaussian Splatting scenes, and label them with depth predictions from the large FoundationStereo model instead of the reconstructed 3D geometry. This gives a cheaper, faster way to build custom stereo training data, and a small stereo network fine-tuned on it matches or beats earlier NeRF-based dataset generation on real-world benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Zero-shot benchmark interpretation hinges on unverified claim that FoundationStereo's training data excludes all test benchmarks; without this, 3DGS+FS is not zero-shot.","rationale":"I agree with the reader's identification of the synthetic-only training claim as the weakest link. The paper's central evaluation is a zero-shot test; the entire 3DGS+FS pipeline feeds FoundationStereo's predictions into training labels, so the premise that the teacher has never seen the test benchmarks is essential. The paper's one-sentence assertion lacks citation or dataset provenance, and the FoundationStereo reference is a preprint, making this precisely the kind of omitted support that should gate a conditional verdict. I considered an alternative concern—the initialization confound in the NeRF-Stereo comparison—but that weakens the 'consistently outperforms' phrasing without invalidating the broader conclusion that 3DGS+FS is competitive on Midd-T and ETH3D. The leakage issue, if true, would invalidate the entire benchmark and the generalization claims derived from it. Since the concern is not yet confirmed, the conditional verdict is appropriate rather than a harsher rejection.","tokens_in":10881,"tokens_out":7187,"duration_ms":81227,"concrete_test":"Verify the training corpus of FoundationStereo (Wen et al., 2025) by inspecting the official paper, repository, and any released dataset manifests: list every dataset used for training, pre-training, and validation. If any of KITTI 2012/2015, Middlebury (any version), or ETH3D appears, repeat the fine-tuning protocol of Section 4 using a teacher that is provably synthetic-only (e.g., a FoundationStereo checkpoint trained before any real-data exposure, or an equivalent model with a published synthetic-only corpus). If the 3DGS+FS margins on Midd-T and ETH3D in Table 1 shrink or reverse, the zero-shot generalization claim is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 asserts: 'It is also worth noting that FoundationStereo was trained purely on synthetic data, eliminating any concerns of data leakage from the test datasets.' This premise is load-bearing because 3DGS+FS uses FoundationStereo's disparity predictions as pseudo-ground-truth to fine-tune RAFT-Stereo. Any exposure of FoundationStereo to KITTI-15, Middlebury v3, or ETH3D—during pre-training, fine-tuning, validation, or via dataset construction—would leak test-set signal into the training labels, turning the zero-shot evaluation into a measure of teacher-memorization rather than dataset-generation quality. The paper provides no citation, code-repo pointer, dataset manifest, or other verification for this claim. A secondary confound is that 3DGS+FS is initialized from the SceneFlow checkpoint while the reproduced NeRF-Stereo is trained from scratch; the claimed 'consistently outperforms' advantage over NeRF-Stereo may therefore reflect initialization rather than dataset merit. Table 1 also shows original NeRF-Stereo winning on KITTI-15 and Midd-T Q, so the contribution's 'consistently outperforms' phrasing is already overstated unless those results are dismissed as unreproducible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a pipeline for synthetic stereo dataset generation based on 3D Gaussian Splatting (3DGS) and expert knowledge transfer. The authors render stereo pairs from 3DGS reconstructions of the NeRF-Stereo scene dataset and use predictions from FoundationStereo as pseudo ground-truth disparity. They fine-tune RAFT-Stereo on these datasets and evaluate zero-shot on KITTI-15, Middlebury v3, and ETH3D. The main findings are that geometry extracted from 3DGS-based reconstruction methods (plain 3DGS, PGSR, 2DGS, GOF) is too noisy to train a competitive stereo network, whereas pseudo-labels from FoundationStereo yield strong zero-shot performance, often exceeding the reproduced NeRF-Stereo and approaching the SceneFlow-pretrained RAFT-Stereo baseline. The paper also reports that about 10% of scenes fail to reconstruct and that the original NeRF-Stereo results are not reproducible.","tokens_in":11163,"tokens_out":9090,"duration_ms":86628,"significance":"The proposed pipeline is a low-cost alternative to synthetic datasets built from explicit 3D assets and to the NeRF-Stereo pipeline, since it avoids the need for ground-truth depth sensors or hand-crafted assets. The negative result about the in-the-wild robustness of state-of-the-art Gaussian splatting meshing methods is useful for the community. The method is timely and clearly described. If the claims are properly qualified and the zero-shot interpretation is verified, the contribution would be a solid empirical study. However, the current version overstates the results and leaves a key assumption about the teacher model's training data unverified.","major_comments":[{"comment":"The statement that 'FoundationStereo was trained purely on synthetic data, eliminating any concerns of data leakage from the test datasets' is load-bearing for the zero-shot evaluation of 3DGS+FS, but no citation, repository link, or training-data manifest is provided to support it. If FoundationStereo has any exposure to KITTI-15, Middlebury, or ETH3D (e.g., through pre-training, fine-tuning, or validation), the pseudo-labels used to train RAFT-Stereo would leak test-set information and the reported zero-shot generalization would be inflated. Please either verify this claim with a precise reference to the FoundationStereo paper (including its training-data section) or remove the 'eliminating' phrasing and qualify the zero-shot interpretation accordingly.","section":"Section 3.4"},{"comment":"The contribution bullet claiming that the expert knowledge transfer setup 'consistently outperforms other methods, including the state-of-the-art NeRF-Stereo method' is not supported by the reported numbers. In Table 1, 3DGS+FS has a larger error than the RAFT-Stereo baseline on KITTI-15 (5.52 vs. 5.46), and the original NeRF-Stereo results are better on KITTI-15 (5.41) and Midd-T Q (8.05 vs. 9.00). The nuanced statement in Section 5 ('outperforms all other methods on the Midd-T and ETH3D datasets, and is close to matching the RAFT-Stereo performance on KITTI-15') is accurate and should be used consistently in the abstract, introduction, and conclusion.","section":"Section 1 (Contributions) and Table 1"},{"comment":"All fine-tuning results are reported from a single run, and no variance or error bars are given. Several of the performance differences that are central to the paper's claims are small (e.g., KITTI-15 All: 5.52 vs. 5.41 for original NeRF-Stereo; ETH3D Noc: 2.14 vs. 2.29 for the RAFT-Stereo baseline). Because the best checkpoint is selected on a validation suite, the reader cannot tell whether these differences are within run-to-run noise. Please report results over multiple random seeds or provide an alternative justification that the comparisons are stable.","section":"Section 4 and Table 1"},{"comment":"The paper reports that about 10% of scenes or meshes could not be reconstructed with 3DGS, GOF, 2DGS, or PGSR, but it does not analyze whether the remaining scenes are systematically biased toward easier cases (e.g., less motion blur, more texture). This potential selection bias could affect the relative comparison between methods, since the dataset used for all methods is drawn from the successfully reconstructed scenes. Please discuss the failure cases in more detail and, if possible, show that the comparisons are robust to the excluded scenes.","section":"Section 6"}],"minor_comments":[{"comment":"The number of scenes and the total number of rendered stereo pairs used for fine-tuning are never stated; please report the dataset size and composition.","section":"Section 4"},{"comment":"The procedure for selecting 'the five cameras with maximized observability' is not specified; please define the criterion and the number of cameras used for each scene.","section":"Section 4"},{"comment":"The row 'NeRF-Stereo α ✓' is not described in Section 4; it is unclear whether this is the authors' fine-tuned version of NeRF-Stereo renderings, and the caption should explain how it was produced.","section":"Table 1"},{"comment":"The statement that spurious faces in the GOF method were 'manually removed' should be disclosed in the main text, as manual post-processing can bias the qualitative comparison.","section":"Figure 2 caption"},{"comment":"The observability-based filtering is described qualitatively; an algorithmic description or pseudo-code would improve reproducibility.","section":"Section 3.3"},{"comment":"Please define the units and coordinate frame of z_depth and note that baseline b and focal length f must be expressed in consistent units.","section":"Equation (2)"},{"comment":"The sentence 'none of them (including the reproduced NeRF-Stereo) are comparable to the RAFT-Stereo baseline' is confusing because 3DGS+FS is comparable or better; please rephrase to refer only to 3DGS, PGSR, and NeRF-Stereo*.","section":"Section 5"},{"comment":"The description of rerunning COLMAP without refining principal points could be clarified, and the authors should state whether this is a standard pre-processing step for 3DGS; otherwise, it may introduce a confound in the comparison with NeRF-Stereo, which used the original poses.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"The paper's claim about FoundationStereo's training data should be checked against the FoundationStereo paper before publication; if the teacher model has indeed seen the test benchmarks, the main experimental contribution collapses. The authors might also be encouraged to verify the reproducibility of the original NeRF-Stereo numbers they dismiss, as this is the basis for their comparison. The scope of the paper (empirical study with a proposed pipeline) fits the venue, but the lack of code/data release is a concern for a reproducibility-focused journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a read: it tests a simple, practical idea—use 3DGS to render stereo pairs, get pseudo-disparity labels from FoundationStereo, and fine-tune RAFT-Stereo on that synthetic data. On the benchmarks it wins, the gains over the reproduced NeRF-Stereo baseline are real and nontrivial, and the observability-based camera selection is a nice heuristic. The negative result on 3DGS mesh geometry—that it is still too noisy for in-the-wild stereo training—is a useful empirical finding, supported by the downstream task evaluation. The paper is also commendably honest about instabilities: COLMAP failures, VRAM limits, and the ~10% of scenes that did not reconstruct.\n\nThe soft spots are in proportion. First, the abstract and contributions claim the method 'consistently outperforms other methods, including the state-of-the-art NeRF-Stereo.' The table does not support that. The original NeRF-Stereo numbers are better on KITTI-15 and Midd-T Q; the authors dismiss those as unreproducible, but that does not turn a mixed result into a consistent win. The claim should be narrowed to what is actually shown. Second, the comparison is not apples-to-apples: 3DGS+FS fine-tunes from the SceneFlow checkpoint, while the reproduced NeRF-Stereo trains from scratch. The advantage may come as much from initialization as from the dataset. An ablation—fine-tuning NeRF-Stereo from SceneFlow, or training 3DGS+FS from scratch—is needed to separate these. Third, the statement that FoundationStereo was trained purely on synthetic data is load-bearing for the zero-shot interpretation, but it is asserted without a citation or verification. If FoundationStereo saw any of the test sets during its own training, the evaluation leaks signal through the teacher. That is fixable with a reference to the FoundationStereo paper or a data manifest, but it has to be fixed. Minor issues: no code or data released, single-run results, and the small N on some benchmarks—none of these alone are fatal, but they add up.\n\nOverall, this is a solid empirical contribution with an overclaimed headline and a confound in the main comparison. A serious referee can sort that out; the underlying pipeline and the negative geometry result are worth preserving. I would send it to peer review, but with a clear request for major revision: fix the claim, add the initialization ablation, and verify the FoundationStereo training-data premise.","headline":"Useful 3DGS+FoundationStereo pipeline with honest reporting, but the 'consistently outperforms' claim is overstated and the comparison to NeRF-Stereo is confounded by initialization.","tokens_in":11636,"tokens_out":4490,"would_cite":true,"duration_ms":50844,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that synthetic stereo training data generated by 3D Gaussian Splatting renderings combined with FoundationStereo pseudo-depth outperforms prior NeRF-based generation methods on zero-shot generalization benchmarks.","keywords":["3D Gaussian Splatting","synthetic stereo dataset","stereo matching","zero-shot generalization","FoundationStereo","RAFT-Stereo","knowledge distillation","novel view synthesis"],"falsifier":"A direct check of FoundationStereo's training distribution would settle the matter: if any of the evaluation datasets (Middlebury v3, ETH3D, KITTI-15) or images derived from them appear in its training data, the zero-shot comparison is invalidated. A complementary test would be to repeat the fine-tuning with a stereo teacher whose synthetic-only training set is verified to exclude the test benchmarks and see whether the gap to NeRF-Stereo persists.","tokens_in":10760,"feed_emoji":"📷","tokens_out":9761,"duration_ms":95120,"temperature":0.7,"pith_summary":"This paper tries to establish that 3D Gaussian Splatting can replace NeRFs as the basis for generating synthetic stereo training data for deep stereo networks. The proposed pipeline renders stereo image pairs from a fitted 3DGS scene and uses disparity predictions from the FoundationStereo model as pseudo-ground-truth labels, rather than relying on the raw reconstructed geometry. Fine-tuning a RAFT-Stereo network on these pairs yields competitive zero-shot performance on Middlebury, ETH3D, and KITTI-15 benchmarks, beating the reproducible NeRF-Stereo baseline on Middlebury and ETH3D while staying close to the SceneFlow-pretrained baseline on KITTI-15. The significance is that high-quality training data can be produced cheaply from freely captured images, without expensive depth sensors or hand-crafted synthetic scenes, and that combining a renderer with a strong pretrained stereo model transfers better than using noisy reconstructed geometry.","feed_headline":"3D splatting plus a stereo AI beats NeRF for synthetic stereo data","feed_subtitle":"Fine-tuning a lightweight stereo model on generated pairs matches or beats prior methods without hand-labeled depth.","key_machinery":"The key machinery is an expert knowledge transfer pipeline with three stages. First, COLMAP estimates camera poses from a free image collection, and a 3D Gaussian Splatting scene is fitted to those images. Second, stereo pairs are rendered from the 3DGS scene at virtual baselines; instead of using the splat geometry or a meshed reconstruction (e.g., PGSR), an observability score computed per camera vertex is used to select the best-viewed poses for the dataset. Third, the rendered left-right image pairs are passed to FoundationStereo, a stereo vision foundation model, to produce dense disparity pseudo-labels, and a lightweight RAFT-Stereo network is fine-tuned on these pairs. The load-bearing identity is the conversion of rendered depth to disparity via $d = f b / z_{\\text{depth}}$, combined with the realization that the pseudo-labels are cleaner than any geometry the 3DGS method itself can extract.","core_discovery":"The central discovery is that the explicit geometry recovered from 3D Gaussian Splatting is too noisy and artifact-laden to serve as ground truth for training a stereo network, but the same rendered images, when paired with disparity pseudo-labels from the FoundationStereo model, produce training data that improves zero-shot generalization. In the comparison of fine-tuned RAFT-Stereo models, the proposed 3DGS+FS setup achieves the best or second-best error on the Middlebury training set (best at Full and Half resolutions, second at Quarter), the best error on ETH3D, and comes close to the SceneFlow-pretrained RAFT-Stereo on KITTI-15. The originally published NeRF-Stereo numbers are lower on KITTI-15 and Middlebury Quarter, but the paper argues those results are not reproducible and therefore not representative, and its own reproduced NeRF-Stereo numbers are worse across the board. The paper attributes the gain to FoundationStereo's ability to infer clean geometry from limited visual signal, and to 3DGS's ability to render photorealistic images despite its poor underlying geometry.","pith_inferences":["The success of the teacher-student pairing suggests the recipe generalizes to other teacher models and other dense prediction tasks (e.g., monocular depth or optical flow), where a photorealistic renderer supplies images and a foundation model supplies labels.","The observability metric could be used as a general camera-selection heuristic in any multi-view reconstruction or capture setting, not just dataset generation, to maximize reconstruction quality per added view.","A multi-teacher ablation (e.g., replacing FoundationStereo with other stereo foundation models) would clarify how much of the gain comes from teacher capability versus the rendering pipeline itself, a distinction the current single-teacher design cannot resolve.","The results imply that benchmark performance on standard 3D reconstruction datasets is not a reliable proxy for a method's usefulness as a pseudo-ground-truth generator; a downstream stereo task could serve as an additional robustness benchmark for surface reconstruction methods."],"forward_implications":["Fine-tuning on 3DGS+FS data improves recovery of thin structures and complex layered depth discontinuities compared with a SceneFlow-only RAFT-Stereo baseline, as shown in the qualitative examples with bicycle cables and flowers.","The reconstructed meshes from PGSR, 2DGS, and GOF, even after observability filtering, are too artifact-heavy to teach a stereo network, implying that visually good rendering does not guarantee usable geometry.","The pipeline enables custom synthetic stereo datasets from freely captured images, removing the need for depth sensors or hand-authored 3D scenes, and is faster than NeRF-based generation due to 3DGS rendering speed.","The expert knowledge transfer setup acts as a form of distillation: a small real-time-capable student (RAFT-Stereo) inherits the zero-shot stereo ability of a much larger teacher (FoundationStereo).","Current 3DGS-based surface reconstruction methods, despite strong benchmark scores, degrade on in-the-wild scenes, which the proxy task of stereo fine-tuning exposes."],"supporting_citations":[{"why":"Provides the 3D Gaussian Splatting representation and rendering equation used to produce the synthetic stereo pairs.","marker":"[15]"},{"why":"Defines the NeRF-Stereo baseline, the dataset and rendering protocol (poses, baselines) adapted here, and the evaluation protocol.","marker":"[29]"},{"why":"Supplies the FoundationStereo model that generates the pseudo-disparity ground truth in the expert knowledge transfer setup.","marker":"[32]"},{"why":"The RAFT-Stereo architecture and fine-tuning procedure used as the student network and training paradigm.","marker":"[18]"},{"why":"The SceneFlow dataset that provides the pretraining for the baseline RAFT-Stereo checkpoint and for the comparative evaluation.","marker":"[19]"},{"why":"Middlebury v3 datasets used both as validation (Midd-A, Midd-21) and as the Midd-T zero-shot test benchmark.","marker":"[24]"},{"why":"ETH3D benchmark used as a zero-shot test set for the trained stereo models.","marker":"[27]"},{"why":"KITTI 2015 dataset used as a zero-shot test set for the trained stereo models.","marker":"[20]"}],"fun_headline_variants":["Splat renders, stereo AI supplies depth: better synthetic training data","3DGS images plus FoundationStereo depth beat NeRF-generated stereo sets","For stereo model training, skip splat geometry, use AI pseudo-labels","3D splatting for images, FoundationStereo for depth: a winning combo","Clean depth from FoundationStereo makes 3DGS synthetic data useful"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire zero-shot claim rests on FoundationStereo having been trained purely on synthetic data, so that its pseudo-labels carry no information from the Middlebury, ETH3D, or KITTI test sets; if that premise is false, the 'zero-shot' generalization becomes partly circular.","fun_headline_variants_meta":{"raw":{"variants":["Splat renders, stereo AI supplies depth: better synthetic training data","3DGS images plus FoundationStereo depth beat NeRF-generated stereo sets","For stereo model training, skip splat geometry, use AI pseudo-labels","3D splatting for images, FoundationStereo for depth: a winning combo","Clean depth from FoundationStereo makes 3DGS synthetic data useful"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1441,"prompt_tokens":974,"completion_tokens":467,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":367}},"tokens_in":590,"tokens_out":467,"duration_ms":5592,"temperature":1.0,"reasoning_tokens":367,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:30:32.651316+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct check of FoundationStereo's training distribution would settle the matter: if any of the evaluation datasets (Middlebury v3, ETH3D, KITTI-15) or images derived from them appear in its training data, the zero-shot comparison is invalidated. A complementary test would be to repeat the fine-tuning with a stereo teacher whose synthetic-only training set is verified to exclude the test benchmarks and see whether the gap to NeRF-Stereo persists.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics (TOG), 42:1 – 14, 2023","cited_arxiv_id":null,"evidence_quote":"Provides the 3D Gaussian Splatting representation and rendering equation used to produce the synthetic stereo pairs."},{"cited_title":"Nerf-supervised deep stereo.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 855–866, 2023","cited_arxiv_id":null,"evidence_quote":"Defines the NeRF-Stereo baseline, the dataset and rendering protocol (poses, baselines) adapted here, and the evaluation protocol."},{"cited_title":"Foundationstereo: Zero- shot stereo matching.arXiv, 2025","cited_arxiv_id":null,"evidence_quote":"Supplies the FoundationStereo model that generates the pseudo-disparity ground truth in the expert knowledge transfer setup."},{"cited_title":"Raft-stereo: Multilevel recurrent field transforms for stereo matching","cited_arxiv_id":null,"evidence_quote":"The RAFT-Stereo architecture and fine-tuning procedure used as the student network and training paradigm."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The SceneFlow dataset that provides the pretraining for the baseline RAFT-Stereo checkpoint and for the comparative evaluation."},{"cited_title":"High-resolution stereo datasets with subpixel-accurate ground truth","cited_arxiv_id":null,"evidence_quote":"Middlebury v3 datasets used both as validation (Midd-A, Midd-21) and as the Midd-T zero-shot test benchmark."},{"cited_title":"Sch ¨onberger, S","cited_arxiv_id":null,"evidence_quote":"ETH3D benchmark used as a zero-shot test set for the trained stereo models."},{"cited_title":"Menze, C","cited_arxiv_id":null,"evidence_quote":"KITTI 2015 dataset used as a zero-shot test set for the trained stereo models."}],"review_version":1}