{"id":"bda41806-7cdb-4de9-b54a-1ef002e92b61","arxiv_id":"2607.08772","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":6,"one_line_summary":"A teacher-student semi-supervised framework adapts VGGT-based feed-forward 3D reconstruction to underwater environments using synthetic degraded on-land data and unlabeled real underwater video, achieving state-of-the-art underwater depth and point cloud estimation without any underwater 3D labels.","lead":"This paper adapts feed-forward 3D reconstruction models to underwater scenes using semi-supervised learning on unlabeled underwater video, requiring no underwater 3D annotations. A smart generalist might read it because underwater 3D perception is critical for marine robotics, archaeology, and seafloor mapping, and this approach removes the annotation bottleneck.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Cross-view consistency loss enforces consistency, not correctness; ablation shows semi-supervised branch is fragile without careful constraints","rationale":"The reader's identified weakest assumption — pseudo-label reliability on degraded underwater video — is the correct load-bearing concern. The ablation in Table 5 provides direct evidence that this concern is real: the semi-supervised branch alone degrades performance, and only becomes beneficial with strong augmentation and the cross-view consistency loss. This means the paper's contribution is more fragile than the headline results suggest — the method works, but the margin between 'helpful semi-supervised learning' and 'confirmation of teacher bias' depends critically on the augmentation and masking strategy. However, the concern does not overturn the verdict because: (1) the final method does improve over baselines on multiple external benchmarks with independent ground truth, making it unlikely that the gains are purely artifacts of pseudo-label circularity; (2) the ablation transparently shows the fragility, which is good scientific practice; (3) the limitations section acknowledges the failure modes. The verdict of ACCEPT with MODERATE confidence is appropriate. The concern would escalate if the proposed concrete test showed that the static mask fails to filter high-error teacher predictions, which would indicate the cross-view loss is reinforcing rather than correcting teacher bias.","tokens_in":18644,"tokens_out":2661,"duration_ms":87733,"concrete_test":"Run the EMA teacher model (at final checkpoint) on Sea-thru and FLSea-Stereo evaluation sets and directly measure pseudo-label depth quality (Rel, δ1) before any student refinement. Then measure what fraction of pixels passing the static mask (Eq. 8, k=N-2) have teacher error below the median. If the teacher's masked-pixel error is comparable to or worse than the unmasked error, the static mask is not filtering unreliable predictions and the cross-view loss is reinforcing teacher bias rather than correcting it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader correctly identifies pseudo-label reliability as the key assumption. The ablation (Table 5) provides direct evidence that this concern is real, not hypothetical: adding unlabeled real underwater video WITHOUT strong augmentation and cross-view loss degrades performance (Sea-thru Rel: 0.173→0.181; FLSea Stereo Rel: 0.135→0.165). The semi-supervised branch only becomes net-positive when combined with strong augmentation and the cross-view consistency loss. This fragility suggests the teacher's pseudo-labels on real underwater video are indeed noisy or systematically biased. The cross-view consistency loss (Eq. 10) and static mask (Eq. 8) mitigate this by selecting depth-consistent pixels across views, but they enforce consistency, not correctness: a teacher that is systematically wrong in a cross-view-consistent way (e.g., uniform scale error on low-texture regions) would pass the static mask and reinforce errors. The supervised loss on synthetic data provides an anchor, but the ratio of unlabeled to labeled samples is 1:3, and the unsupervised loss weight ramps to 0.5, so the noisy branch has substantial influence. That said, the consistent improvements on external benchmarks with independent ground truth (Sea-thru, FLSea, SQUID — 10-20%+ gains) are strong evidence that the net effect is positive despite this fragility. The concern lands but does not overturn the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The manuscript proposes Wat3R, a cross-domain semi-supervised framework for adapting feed-forward 3D reconstruction models (specifically VGGT) to underwater environments without requiring any underwater 3D annotations. The method follows a Mean Teacher paradigm: synthetic underwater degradation is applied to labeled on-land data to initialize geometric priors, while unlabeled real underwater video is used for semi-supervised refinement. A cross-view consistency loss with a static mask is introduced to mitigate the impact of degraded views by aggregating geometric cues across multiple views. The authors also construct Water3D, a new multi-view underwater dataset with 42 scenes featuring camera pose and depth annotations. The method is evaluated on multi-view depth estimation, point cloud estimation, camera pose estimation, and monocular depth estimation, demonstrating consistent improvements over VGGT and other recent feed-forward baselines.","tokens_in":18857,"tokens_out":703,"duration_ms":200873,"significance":"The paper addresses a practical and significant problem: the domain gap between terrestrial and underwater 3D vision. The use of unlabeled video to bypass the annotation bottleneck is well-motivated. The authors provide public code and a new benchmark dataset (Water3D), which adds considerable value to the community. The experimental evaluation is comprehensive, spanning multiple public datasets and tasks. The ablation studies transparently show the contribution of each component, including the finding that the semi-supervised branch requires careful constraints to be net-positive. The cross-view consistency loss is a reasonable mechanism to handle view-dependent degradation.","major_comments":[],"minor_comments":[{"comment":"§3.3, Eq. (8): The static mask threshold is set to k=N-2. The text states this is a 'conservative setting.' It would help to briefly justify this specific choice or provide a sensitivity analysis, as the mask is load-bearing for the cross-view loss.","section":null},{"comment":"Table 1: The shaded rows indicating two-stage pipelines (UIE + VGGT) are mentioned in the caption, but the visual distinction (shading) is not easily discernible in standard text review. Ensure the formatting is clear in the final version.","section":null},{"comment":"§4.1: The ratio of unlabeled to labeled samples is 1:3, and the unsupervised loss weight ramps to 0.5. It would be beneficial to include a brief discussion or ablation on how sensitive the final performance is to this specific ratio and peak weight.","section":null},{"comment":"Appendix D.2, Table A3: The comparison with COLMAP is useful. However, the 'Failure rate' for COLMAP+MINIMA is 99/152, which is very high. It would be helpful to clarify whether the reported metrics for COLMAP are computed only on the successfully reconstructed scenes, which would make the direct comparison with Wat3R (0 failure rate) less straightforward.","section":null},{"comment":"Fig. 1: The y-axis labels of the bar charts in the top right are rotated vertically and are somewhat difficult to read. Consider adjusting the layout for clarity.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The skeptic's concern regarding the reliability of Mean Teacher pseudo-labels is valid, but the manuscript's own ablation (Table 5) transparently addresses this by showing that the semi-supervised branch is only beneficial when combined with strong augmentation and the cross-view loss. The consistent gains on external benchmarks with independent ground truth suggest the net effect is positive. The central claim is sound and well-supported."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The headline: this paper adapts VGGT (a feed-forward multi-view geometry model) to underwater scenes using a Mean Teacher semi-supervised framework, requiring zero underwater 3D annotations. The results are consistently strong — 10-20%+ relative gains over VGGT and other SOTA methods across Sea-thru, FLSea, SQUID, and their new Water3D benchmark. Code and data are promised. This is a practical contribution that solves a real bottleneck for underwater vision. The cross-view consistency loss with static masking is a reasonable domain-specific mechanism: it backprojects teacher depth across views, checks for depth consistency, and only supervises on pixels that survive the filter. The Water3D dataset itself is a useful contribution — 42 scenes with pose and depth annotations, filling a gap in underwater benchmarks. The ablation in Table 5 is honest and revealing. Adding unlabeled real video without strong augmentation and cross-view loss actually degrades performance (Sea-thru Rel goes from 0.173 to 0.181; FLSea Stereo from 0.135 to 0.165). The semi-supervised branch only becomes net-positive when combined with strong augmentation and the cross-view consistency loss. This confirms the stress-test concern: the teacher's pseudo-labels on degraded underwater video are noisy, and the framework is fragile without the right constraints. The cross-view loss enforces consistency, not correctness — a teacher that is systematically wrong in a cross-view-consistent way would still pass the static mask. That said, the supervised loss on synthetic data anchors the model, and the consistent gains on external benchmarks with independent ground truth are strong evidence the net effect is positive. The Water3D benchmark has selection bias — only COLMAP-successful scenes are retained — but the central claim doesn't rest solely on Water3D since external benchmarks show the same improvements. The framework also generalizes beyond VGGT: applying the same approach to π3 (Table A5) yields similar gains, which is a good sign. Overall, this is a solid, well-executed paper with a practical contribution. The weakest point is the fragility of the semi-supervised branch, but the authors acknowledge this and the experimental evidence supports the central claim. It deserves a serious referee.","headline":"Semi-supervised adaptation of feed-forward 3D reconstruction to underwater scenes without annotations — works well, but the semi-supervised branch is fragile without careful constraints.","tokens_in":19639,"tokens_out":534,"would_cite":true,"duration_ms":58419,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Underwater 3D Reconstruction Without Any Underwater Annotations","keywords":[],"falsifier":"If the teacher network produces systematically biased geometry on real underwater video—particularly in green-water or high-turbidity conditions where the pretrained model has no prior—the student would reinforce those errors. The cross-view static mask filters per-pixel inconsistency but cannot detect systematic biases that are consistent across all views of a degraded scene. The paper acknowledges this limitation for open-water and very deep scenes, but the boundary between where the method works and where it fails is not sharply characterized.","tokens_in":18873,"feed_emoji":"🌊","tokens_out":1123,"duration_ms":131339,"temperature":0.7,"pith_summary":"The paper claims that a feed-forward 3D geometry model trained on land can be adapted to underwater scenes without a single annotated underwater depth map, camera pose, or point cloud. The mechanism is a teacher-student semi-supervised loop: synthetic underwater degradation is applied to existing on-land 3D datasets to seed geometric priors, then a teacher network generates pseudo-labels on real unlabeled underwater video while a student network learns from both. A cross-view consistency loss reprojects depth across views and supervises only on pixels that are geometrically stable across multiple viewpoints, filtering out dynamic objects and regions where scattering has destroyed reliable structure. The authors also construct Water3D, a 42-scene underwater benchmark with pose and depth annotations, and show that their adapted model outperforms current feed-forward baselines on multi-view depth, point cloud reconstruction, camera pose, and monocular depth tasks across five public datasets plus their own.","feed_headline":"Underwater 3D Reconstruction Without Any Underwater Annotations","feed_subtitle":"A semi-supervised teacher-student loop adapts land-trained geometry models to underwater scenes using only unlabeled video, outperforming S0","key_machinery":"Mean Teacher semi-supervised framework with EMA-updated teacher producing pseudo-labels on unlabeled underwater video; physics-based underwater image formation model (direct transmission attenuation, backscatter, background light) applied to on-land 3D-annotated datasets to create labeled synthetic underwater training data; cross-view geometry consistency loss that backprojects teacher depth from one view, reprojects into other views, and supervises student depth only on pixels passing a multi-view static mask (depth-consistent foreground across N-2 views); sequence-level augmentation with frame shuffling and rotation to prevent pose collapse; VGGT as base feed-forward geometry backbone.","core_discovery":"The central claim is that the domain gap between air-trained 3D geometry models and underwater scenes can be bridged entirely through synthetic degradation plus semi-supervised consistency training on unlabeled video, without any underwater 3D ground truth. The cross-view consistency loss is the key technical device: it compensates for information lost to attenuation and scattering in one view by pulling geometric evidence from other views of the same scene, applying supervision only where multiple views agree on depth. This mechanism, combined with a physics-based underwater image formation model for synthetic data generation, allows a large pretrained geometry transformer to generalize to ","pith_inferences":["The reliance on a static mask that becomes sparse in open-water or highly turbid scenes suggests a natural performance ceiling: the method may work well in structured underwater environments (reefs, shipwrecks, seabed with texture) but degrade precisely where underwater robotics needs it most—featureless open water with poor visibility.","The fact that the same framework improves both multi-view and monocular depth estimation suggests the cross-view training is teaching the model something general about separating underwater degradation from scene geometry, not just memorizing multi-view correspondences. This raises the question of whether the learned representations could transfer to other degraded-media tasks like imaging through","If the teacher's pseudo-labels are systematically biased in certain underwater conditions—for example, consistently overestimating depth in green-water scenes—the student would reinforce that bias. The cross-view mask filters individual pixel errors but may not catch systematic biases that are consistent across views."],"forward_implications":["If the approach generalizes, any domain with abundant unlabeled video but no 3D annotations—subterranean, foggy outdoor, or endoscopic scenes—could be adapted similarly by combining synthetic degradation with cross-view consistency training.","The finding that underwater image enhancement as a preprocessing step does not improve 3D geometry estimation suggests the field should shift from restoration-then-reconstruction pipelines toward joint geometry-aware adaptation.","The Water3D benchmark with 42 scenes and both depth and pose annotations fills a gap that could recalibrate how underwater 3D methods are evaluated, since prior datasets lacked comprehensive geometric annotations.","The cross-view static masking strategy, which naturally sparsifies when scenes are dynamic or turbid, provides a principled mechanism for self-limiting supervision that could apply to any semi-supervised multi-view learning setting with unreliable regions."],"fun_headline_variants":["Underwater 3D Geometry From Unlabeled Video Alone","No Underwater Labels Needed for 3D Geometry Reconstruction","Bridging Air-Trained 3D Models to Underwater Scenes Without Labels","Cross-View Consistency Replaces Underwater 3D Annotations","Adapting Land-Trained Geometry Models to Underwater Video"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The teacher network's depth and pose predictions on real underwater video are accurate enough to serve as training targets for the student. The entire semi-supervised branch depends on this; if the teacher is systematically wrong on turbid or low-visibility regions, the student learns to reproduce those errors. The cross-view consistency mask is designed to filter unreliable pixels, but it assumes enough views have overlapping static structure, which the authors acknowledge稀疏","fun_headline_variants_meta":{"raw":{"variants":["Underwater 3D Geometry From Unlabeled Video Alone","No Underwater Labels Needed for 3D Geometry Reconstruction","Bridging Air-Trained 3D Models to Underwater Scenes Without Labels","Cross-View Consistency Replaces Underwater 3D Annotations","Adapting Land-Trained Geometry Models to Underwater Video","Underwater Depth Estimation Without Any Annotated Data","Semi-Supervised Loop Adapts 3D Models to Underwater Scenes","Unlabeled Underwater Video Suffices for 3D Geometry Learning","From Air to Water: 3D Geometry Without Underwater Labels","Teacher-Student Framework Learns Underwater Geometry Label-Free"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":1291,"prompt_tokens":514,"completion_tokens":777,"prompt_tokens_details":null},"tokens_in":514,"tokens_out":777,"duration_ms":39455,"temperature":1.0,"reasoning_tokens":603,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T01:26:16.718561+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If the teacher network produces systematically biased geometry on real underwater video—particularly in green-water or high-turbidity conditions where the pretrained model has no prior—the student would reinforce those errors. The cross-view static mask filters per-pixel inconsistency but cannot detect systematic biases that are consistent across all views of a degraded scene. The paper acknowledges this limitation for open-water and very deep scenes, but the boundary between where the method works and where it fails is not sharply characterized.","supporting_citations":[],"review_version":1}