{"id":"a0e39c70-c242-4779-a5ee-ef84097f5038","arxiv_id":"2608.10708","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Self-Geometry adapts frozen 3D vision foundation models per scene by enforcing epipolar and reprojection consistency on pseudo-correspondences from an off-the-shelf matcher, improving pose and depth accuracy.","lead":"This paper introduces a test-time adaptation method that improves the camera poses and 3D geometry produced by vision foundation models, using only 2D pixel matches between images as supervision, without ground-truth labels. It works as a plug-in for several existing models and takes just a few minutes per scene.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains rely on selecting the TTA checkpoint that minimizes the adaptation loss itself; without a fixed selection rule or per-iteration GT curves, the claimed consistent improvements are not yet established.","rationale":"The reader identified LightGlue correspondence quality as the weakest assumption. That is a real limitation and is acknowledged in S.VI, but it does not by itself undermine the central claim: the filtering and robust losses can tolerate imperfect correspondences, and the tables show gains despite low raw precision. The more decisive, less acknowledged issue is the checkpoint selection protocol. Since the selected checkpoint minimizes the very loss being optimized, the reported numbers are the best over 50 iterations by the adaptation objective; without per-iteration GT or an error bar, one cannot tell whether this selection masks instability or overfitting. This concern appears in the reader's rationale ('checkpoint selection based on the training loss'), so agreement is partial even though the reader's formal weakest_assumption points elsewhere. The proposed check is cheap and would settle whether the central claim holds under a fixed, deployable selection rule, while also addressing the lack of error bars flagged by the reader.","tokens_in":25114,"tokens_out":8709,"duration_ms":79121,"concrete_test":"Re-run Self-Geometry for VGGT on ETH3D and DA3-Base on HiRoom, recording pose and geometry metrics at every TTA iteration (or at iterations 1, 10, 20, 30, 40, 50). Compare three selection rules: the best-self-supervised checkpoint (current paper), the final checkpoint at t=50, and the checkpoint with the best GT metrics as an upper bound. If the final-checkpoint rule yields mean improvements over the frozen baseline that are not positive, or if the best-self-supervised checkpoint's GT gains are substantially below the best-GT checkpoint's gains, the reported 'consistent improvements' depend on the training-loss selection rule and the central claim should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Tables II and III are produced by selecting, per scene, the TTA checkpoint that minimizes the self-supervised objective sqrt(L_ec * L_mvc) over 50 iterations (Algorithm 1, lines 15-17; Sec. IV-A). Because L_ec and L_mvc are computed from the same noisy LightGlue pseudo-correspondences that drive adaptation, the selected checkpoint is the best of 50 iterations with respect to the training signal itself. The paper reports no per-iteration GT curves, no fixed-iteration baseline, and no error bars, so it is unknown whether the best self-supervised checkpoint also has the best GT pose/geometry error, or whether it is simply the iteration most overfit to match noise. If the latter, the 'consistent improvements' central to the abstract may be an artifact of this selection rule rather than a property of the adapted model. This is load-bearing because a deployed pipeline needs a concrete checkpoint selection rule, and the only evidence offered for 'consistent improvements' uses the training-loss-based selection rule.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Self-Geometry, a test-time adaptation (TTA) pipeline for pretrained multi-view 3D vision foundation models (VFMs). The method extracts 2D pixel correspondences with LightGlue and uses them as pseudo ground truth for explicit geometric losses: a point-to-point Multi-View Consistency (MVC) loss on reprojection and a point-to-line Epipolar Consistency (EC) loss based on the Sampson distance. Gradient Disentanglement projects the MVC gradient away from the EC gradient to reduce conflict, Frame Angular-Neighbor (FAN) selects target/source views using SO(3) geodesic distances, and lightweight LoRA adapters are updated per scene. The authors report pose and geometry improvements over frozen VFMs and two TTA baselines across six VFMs (VGGT, π3, DA3-Giant/Large/Base/Small) and four datasets (7Scenes, ETH3D, ScanNet++, HiRoom).","tokens_in":25349,"tokens_out":4185,"duration_ms":44615,"significance":"If the reported gains are robust, Self-Geometry would be a practically valuable, GT-free, teacher-free, architecture-agnostic test-time adaptation recipe, and the explicit-use of correspondences as pseudo-GT is a well-motivated departure from implicit self-consistency methods. The paper is commendable for its clear component-wise ablations (Tables IV–VI), a complexity analysis (Table VII), and for evaluating across a broad set of VFMs and benchmarks. However, the central empirical claim of 'consistent improvements' currently rests on a checkpoint-selection rule that is itself based on the training objective, and the small effect sizes in many settings make this a load-bearing issue rather than a presentation concern.","major_comments":[{"comment":"The reported results select, per scene, the TTA checkpoint that minimizes the self-supervised metric sqrt(L_ec * L_mvc), i.e., the best of 50 iterations with respect to the very objective being optimized. Because L_ec and L_mvc are computed from the same LightGlue pseudo-correspondences that drive adaptation, this is a form of self-selection that can artifactually favor the iteration most overfit to match noise. The paper provides no per-iteration ground-truth curves and no evaluation at a fixed iteration count, so it is unknown whether the best-training-loss checkpoint is also the best ground-truth checkpoint. This is load-bearing for the abstract claim of 'consistent improvements,' especially where mean gains are small (e.g., DA3-Giant-Ours Mean AUC@3 +1.5% in Table II and Mean geometry w/o p. +0.9% in Table III). Please report results with a fixed iteration rule (e.g., iteration 50) alongside the selection-based results, and provide per-iteration GT curves or a scatter plot of training-loss vs. GT error across iterations; error bars across scenes are also needed.","section":"§IV-A and Algorithm 1 (lines 15–17)"},{"comment":"There is a direct inconsistency between the method description and the algorithm. The text states that filtering is applied 'before every TTA iteration' using the model's current predictions, but Algorithm 1 applies filtering only once at initialization (line 2), before the loop begins. These two readings have different implications: per-iteration filtering creates a feedback loop in which the adapted model's own (possibly erroneous) poses/depths prune the supervision set, which could reinforce drift; one-time filtering keeps the pseudo-GT fixed during adaptation. The paper must clarify which implementation is actually used. If filtering is per-iteration, report the sensitivity to filter thresholds and demonstrate that the adapted model's predictions do not progressively bias the correspondence set; if it is one-time, correct the text in §III-B.","section":"§III-B 'Pseudo-Correspondence Filtering' vs. Algorithm 1"},{"comment":"The claim of 'consistent improvements' is weakened by several negative entries that the text acknowledges but does not statistically assess: for VGGT-Ours, AUC@3 drops by 2.9% on ScanNet++ and 8.2% on HiRoom (Table II), and geometry F1 w/o p. drops by 10.6% and 7.5% (Table III); DA3-Giant-Ours and DA3-Large-Ours show near-zero or negative changes on several columns. The authors argue these are 'small absolute drops,' but without per-scene breakdowns, confidence intervals, or significance tests, it is not possible to distinguish genuine degradation from noise or to verify that the positive means are robust. Please supply per-scene results and a statistical summary (e.g., paired differences with confidence intervals or a sign-test across scenes).","section":"Tables II and III"}],"minor_comments":[{"comment":"The projection in Eq. (4) divides by ||∇Lec||^2; if ∇Lec is exactly zero, the update is undefined. Add a small epsilon or a guard for the zero-gradient case.","section":"Eq. (4)"},{"comment":"The text says every hyperparameter is used identically across all VFMs, but for π3 LoRA is inserted only into the encoder. Please clarify whether the 'identical settings' claim refers only to optimization hyperparameters and not to the adapter configuration.","section":"Table S.1 and §S.V"},{"comment":"The abstract and contributions state 'within two minutes per-scene,' but Table VII reports ETH3D totals of 3.29 min for VGGT and 5.36 min on 7Scenes for VGGT, with several other entries above two minutes. Please qualify the timing claim to the specific configuration for which it holds, or report the full distribution.","section":"Abstract and Table VII"},{"comment":"The caption of Figure 4 contains a garbled string of Unicode escape-like tokens ('/uni00000150/...'), which appears to be a rendering artifact; it should be replaced with the intended scene identifiers or removed.","section":"Figure 4 caption"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about checkpoint selection is real and central: with many relative gains in the 1–5% range, selecting the best of 50 iterations by the training loss could account for a substantial portion of the reported improvements. I would like to see a fixed-iteration evaluation and per-scene error bars before accepting. The paper is otherwise well-structured and the method is a reasonable contribution to the TTA literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this paper has a real design idea — explicit epipolar and reprojection losses with gradient disentanglement, applied to frozen VFMs via LoRA at test time — but the main quantitative claim is weaker than it looks. The authors select, per scene, the TTA checkpoint that minimizes sqrt(L_ec * L_mvc), a loss computed from the same LightGlue pseudo-correspondences that drive adaptation. Nothing in the paper shows that this self-supervised best is also the GT best. Without per-iteration GT curves or a fixed-iteration baseline, the reported gains could be partly an artifact of overfitting to match noise. That is load-bearing, not a nitpick.\n\nThe contribution itself is solid. The pose-depth ambiguity argument is clear: the reprojection loss alone cannot separate pose and depth, so they add a depth-independent epipolar loss and project the MVC gradient onto the orthogonal complement of the EC gradient. Ablations confirm that both losses matter and that gradient disentanglement helps. The SO(3)-based view sampler is reasonable and well ablated, and the fixed hyperparameters across six models and four datasets make the recipe practical. The paper is also honest about matcher dependence: raw correspondences hurt in their ablation, and the proposed filtering recovers most of the loss, which is good evidence the supervision pipeline works. Code is only promised, not released.\n\nSoft spots, in order. First, the checkpoint selection issue above. It directly affects every positive percentage in Tables II and III. Second, \"consistent improvements\" is an overstatement; several entries are negative (VGGT on ScanNet++ and HiRoom, for example), even if means are positive. The prose acknowledges this, but the abstract still overclaims. Third, there are no error bars or significance tests, which matters more because the selection rule favors the training loss. These are fixable, but they change the strength of the claim.\n\nWho is this for? Anyone working on test-time adaptation of 3D VFMs or on making feed-forward reconstruction more geometric. The method deserves serious engagement; the evaluation just needs tightening. I would send it to peer review, with the checkpoint selection concern as the main request: add fixed-iteration and per-iteration GT curves, plus error bars, before acceptance.","headline":"Genuinely new TTA recipe for 3D VFMs, but the headline numbers rest on selecting the checkpoint that minimizes the adaptation loss itself, so the 'consistent improvements' claim is not yet established.","tokens_in":25896,"tokens_out":2832,"would_cite":true,"duration_ms":91852,"reading_group":"yes","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 pretrained 3D vision foundation models can be made geometrically consistent at test time, without any ground truth, by using 2D pixel correspondences as pseudo ground truth, and that this adaptation improves both…","keywords":["test-time adaptation","vision foundation models","3D reconstruction","epipolar geometry","multi-view consistency","pseudo-correspondences","LoRA","camera pose estimation"],"falsifier":"Run Self-Geometry on a benchmark of textureless or repetitive-texture scenes with known camera poses, measure LightGlue correspondence precision against ground truth, and check whether pose AUC and geometry F1 improvements vanish when precision falls below a threshold; alternatively, replace the LightGlue correspondences with random matches of the same count—if adaptation still improves, the pseudo-correspondences are not the cause, and if it degrades sharply, the claim's reliance on matcher quality is confirmed.","tokens_in":24932,"feed_emoji":"📐","tokens_out":3894,"duration_ms":38612,"temperature":0.7,"pith_summary":"This paper claims that the multi-view geometric inconsistencies of pretrained 3D vision foundation models can be corrected at test time, without ground truth, by treating 2D pixel correspondences from an external feature matcher as pseudo ground truth. It proposes Self-Geometry, a plug-and-play test-time adaptation pipeline that imposes explicit multi-view geometric constraints on a frozen model through two complementary losses: a point-to-point reprojection loss and a depth-independent epipolar loss, with their gradients disentangled to avoid conflict. Per-scene adaptation is kept light by sampling views with SO(3) geodesic distances and updating only LoRA adapters, finishing in under two minutes on a single GPU. The paper reports consistent gains in pose and geometry estimation over frozen baselines and over prior test-time adaptation methods across six vision foundation models and four benchmark datasets.","feed_headline":"2D matches as pseudo-GT fix 3D vision models at test time","feed_subtitle":"It adapts VGGT, pi3, and Depth Anything 3 per scene, improving pose and geometry across four benchmarks.","key_machinery":"The central machinery is Geometric Disentanglement Optimization, which pairs a point-to-point Multi-View Consistency Loss (reprojection error against pseudo-correspondences) with a point-to-line Epipolar Consistency Loss (Sampson distance against predicted fundamental matrices) and applies Gradient Disentanglement—projecting the Multi-View gradient onto the orthogonal complement of the Epipolar gradient—so the two losses refine pose and depth complementarily. Frame Angular-Neighbor sampling uses SO(3) geodesic distances to select a target view with maximum angular-bin entropy and to sample source views uniformly across bins, and Lightweight Test-Time Adaptation updates only LoRA parameters inserted into the attention blocks' QKV weights, keeping the pretrained model frozen and enabling per-scene adaptation within two minutes.","core_discovery":"Self-Geometry establishes that 2D pixel correspondences—extracted once by LightGlue at scene initialization—provide explicit multi-view geometric supervision that can be imposed on any pretrained vision foundation model at test time without ground-truth labels or a teacher model. The Multi-View Consistency Loss supervises both camera pose and depth through point-to-point reprojection, while the Epipolar Consistency Loss supervises pose alone through a point-to-line Sampson distance, resolving the pose-depth ambiguity. Because both losses act on pose, the method projects the Multi-View Consistency gradient onto the orthogonal complement of the Epipolar Consistency gradient, preventing the gradient conflict that otherwise appears in 42.4% of iterations on the ETH3D scenes tested. Frame Angular-Neighbor sampling selects a target view with maximal SO(3)-bin entropy and samples source views across angular bins, while LoRA adapters inserted into QKV attention weights keep the model mostly frozen. The combination yields consistent Mean improvements in pose AUC and geometry F1 over six VFMs—VGGT, pi3, and four Depth Anything 3 variants—and four benchmarks, with the largest relative gains on wide-baseline outdoor scenes and on smaller, under-fitted models.","pith_inferences":["The method's dependence on matcher quality suggests a natural stress test: on textureless or repetitive-texture scenes where LightGlue precision drops, improvement should shrink; if it does not, the losses are not actually driven by the pseudo-correspondences, which would indicate a different mechanism is at work.","The gradient-disentanglement idea is not specific to epipolar versus reprojection losses; it could transfer to other multi-task test-time settings where two losses supervise overlapping parameters with conflicting gradients, though the paper does not claim this.","The success of per-scene adaptation with explicit geometric losses hints that the same supervision could be added directly to VFM pretraining on pseudo-correspondences from unlabeled video, potentially making test-time adaptation unnecessary for the most common failure modes.","Because the largest relative gains appear on small, under-fitted models, the method could also serve as a diagnostic signal for where a foundation model's geometric prior is weak, guiding where more pretraining data or capacity is needed."],"forward_implications":["Frozen 3D vision foundation models can be specialized to a target scene using only image pairs and a matcher, with no ground-truth poses, depths, or teacher networks, which removes the main data bottleneck for scene-level adaptation.","Explicit geometric constraints outperform implicit self-consistency from model outputs, so future test-time adaptation methods should consider directly supervising with epipolar and reprojection errors rather than distilling predictions.","The gains concentrate where the pretrained model is weakest: wide-baseline outdoor scenes and small under-fitted model variants, suggesting the method is a practical way to extend the useful range of smaller deployable models.","The pipeline is architecture-agnostic, working on VGGT, pi3, and Depth Anything 3 variants without per-backbone redesign, so it can be attached to newly released geometric foundation models as a plug-in.","Per-scene adaptation on a single GPU within roughly two minutes makes the method usable in practical reconstruction pipelines, where it competes with per-scene optimization costs at a fraction of the compute."],"supporting_citations":[{"why":"LightGlue supplies the 2D pixel correspondences that serve as pseudo ground truth for both losses; without it there is no supervision signal.","marker":"[16]"},{"why":"Hartley and Zisserman provide the epipolar constraint and Sampson distance that define the Epipolar Consistency Loss.","marker":"[7]"},{"why":"Gradient surgery supplies the projection operation used to disentangle the Multi-View Consistency gradient from the Epipolar Consistency gradient.","marker":"[20]"},{"why":"LoRA provides the low-rank adapter that is inserted into QKV attention weights, enabling parameter-efficient test-time adaptation.","marker":"[22]"},{"why":"Free-Geometry is the implicit self-consistency test-time adaptation baseline that Self-Geometry argues is outperformed by explicit geometric constraints.","marker":"[12]"},{"why":"VGGT is one of the six pretrained foundation models adapted, anchoring the claim that the method generalizes across architectures.","marker":"[1]"},{"why":"Depth Anything 3 provides four of the six adapted models and the evaluation protocol, including AUC thresholds and F1 pose/unposed geometry metrics.","marker":"[3]"},{"why":"TCO is the prior-based test-time adaptation baseline compared against, using pretrained-predicted poses as its auxiliary prior.","marker":"[36]"}],"fun_headline_variants":["2D matches as pseudo-GT align 3D vision models on the fly","No labels, no teacher: 2D correspondences fix 3D geometry at test time","Self-Geometry: GT-free TTA for 3D vision using 2D matches","Test-time adaptation with 2D pseudo-labels boosts 3D consistency","Pose and geometry improve at test time with 2D pseudo-GT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The external feature matcher's 2D pixel correspondences are accurate and dense enough to serve as ground truth for geometric supervision; if the matcher returns sparse or wrong matches on textureless, repetitive, or wide-baseline scenes, both the epipolar and reprojection losses give weak or misleading gradients and the filtering step can reinforce the model's own errors.","fun_headline_variants_meta":{"raw":{"variants":["2D matches as pseudo-GT align 3D vision models on the fly","No labels, no teacher: 2D correspondences fix 3D geometry at test time","Self-Geometry: GT-free TTA for 3D vision using 2D matches","Test-time adaptation with 2D pseudo-labels boosts 3D consistency","Pose and geometry improve at test time with 2D pseudo-GT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000855,"raw_usage":{"total_tokens":3777,"prompt_tokens":1069,"completion_tokens":2708,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":685,"completion_tokens_details":{"reasoning_tokens":2599}},"tokens_in":685,"tokens_out":2708,"duration_ms":18600,"temperature":1.0,"reasoning_tokens":2599,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:53:54.857001+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Self-Geometry on a benchmark of textureless or repetitive-texture scenes with known camera poses, measure LightGlue correspondence precision against ground truth, and check whether pose AUC and geometry F1 improvements vanish when precision falls below a threshold; alternatively, replace the LightGlue correspondences with random matches of the same count—if adaptation still improves, the pseudo-correspondences are not the cause, and if it degrades sharply, the claim's reliance on matcher quality is confirmed.","supporting_citations":[{"cited_title":"LightGlue: Local feature matching at light speed,","cited_arxiv_id":null,"evidence_quote":"LightGlue supplies the 2D pixel correspondences that serve as pseudo ground truth for both losses; without it there is no supervision signal."},{"cited_title":"Hartley and A","cited_arxiv_id":null,"evidence_quote":"Hartley and Zisserman provide the epipolar constraint and Sampson distance that define the Epipolar Consistency Loss."},{"cited_title":"Free Geometry: Refining 3D Reconstruction from Longer Versions of Itself","cited_arxiv_id":"2604.14048","evidence_quote":"Free-Geometry is the implicit self-consistency test-time adaptation baseline that Self-Geometry argues is outperformed by explicit geometric constraints."},{"cited_title":"VGGT: Visual geometry grounded transformer,","cited_arxiv_id":null,"evidence_quote":"VGGT is one of the six pretrained foundation models adapted, anchoring the claim that the method generalizes across architectures."},{"cited_title":"Depth anything 3: Recovering the visual space from any views,","cited_arxiv_id":null,"evidence_quote":"Depth Anything 3 provides four of the six adapted models and the evaluation protocol, including AUC thresholds and F1 pose/unposed geometry metrics."},{"cited_title":"Learning 3D reconstruction with priors in test time,","cited_arxiv_id":null,"evidence_quote":"TCO is the prior-based test-time adaptation baseline compared against, using pretrained-predicted poses as its auxiliary prior."}],"review_version":1}