{"id":"6eb6f57e-a263-4e78-9f1b-fa21409e965d","arxiv_id":"2608.09590","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A training-only regularizer makes cross-modal 2D-3D matching representations recoverable by weak probes, improving robustness and pose accuracy on 7-Scenes and RGB-D Scenes V2.","lead":"A new training method called TeaMatch uses weak 'probe' models to test whether 2D-3D matching representations survive degradation, then optimizes the representations to be easier for the probes to recover. This improves matching accuracy and pose estimation on two indoor benchmarks with no extra inference cost, but the gains are modest and the core metric is partly self-referential.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TeaMatch's teachability objective is validated only against the teacher's own soft labels; the reported gains may come from linear decodability rather than geometric robustness.","rationale":"TeaMatch is a training-only regularizer; if it works, it should make pair representations robust and geometrically correct. The load-bearing condition is that optimizing probe recoverability of teacher scores transfers to correspondence quality and pose estimation. That condition is least secure because the main teachability signal is self-referential: Eq. (11) uses the teacher's logits as the target, and Eq. (15) measures recoverability of those same logits. The paper's own ablation in Tab. 4 makes the issue salient: the correspondence loss alone changes RR from 75.8 to 79.6 before any GT-anchored losses are added, so the self-referential term carries substantial weight. The inlierness and pose losses are useful but they are standard geometric supervision and do not establish the teachability criterion as defined. I agree with the Reader that this is the weakest assumption. A GT-anchored control is directly implementable: swap y_j for q_j and measure whether the gains survive. If the gains survive, the teacher-score target is dispensable, weakening the claim that teachability, as defined, is the mechanism. If the gains do not survive, the criterion may still be useful, but a second control, probing GT-label recoverability, is needed to show it tracks geometry. This concern does not overturn the Reader's CONDITIONAL verdict; it sharpens the validation that would be required before the central claim can be taken at face value.","tokens_in":16288,"tokens_out":5700,"duration_ms":54393,"concrete_test":"On 7-Scenes, rerun TeaMatch training with the only change that the probe targets in Eqs. (11)-(15) are replaced by GT-anchored soft labels, e.g., q_j = exp(-r_j^2/delta^2) from Eq. (16) or binarized inlier labels at the 5 cm threshold, while keeping L_inlier, L_pose, and all hyperparameters fixed. Compare IR, FMR, and RR against the published TeaMatch and the 2D3D-MATR baseline. If the GT-targeted variant matches or exceeds the published results, the teacher-score target is not necessary and the 'teachability' claim is unsupported; if it performs substantially worse, the self-referential target is essential, and a further check is needed: train the same probes to predict GT inlier labels on baseline versus TeaMatch representations and compare meta-split accuracy to confirm the criterion tracks geometry.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that TeaMatch's teachability criterion improves robust correspondence learning rests on an unvalidated, self-referential proxy. In Eq. (11), the weak probes are trained to recover the teacher's own softened score y_j = sigma(s_j/tau), and the correspondence loss in Eq. (15) measures recoverability of those same scores on the meta split. The paper never compares this target against ground-truth correspondence labels or geometric inlierness, so 'teachable' may simply mean 'linearly decodable in the teacher's current score geometry.' A representation can become easier for a probe to reconstruct without becoming more geometrically correct. The auxiliary inlierness and pose losses (Eqs. 19 and 21) inject GT geometry, but they are secondary and do not validate the teachability criterion itself. The strongest empirical evidence in Tab. 4 makes this concern concrete: the correspondence loss alone accounts for a large part of the RR gain (75.8 to 79.6), before the GT-anchored losses are added. Without a GT-anchored control, the reported improvements over 2D3D-MATR may reflect a linear-decodability regularizer rather than the robust, geometry-preserving representations claimed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TeaMatch, a training-only framework for 2D-3D matching that adds a 'teachability' regularizer to a detection-free coarse-to-fine matcher. The teacher produces pair-level representations for candidate correspondences; during training, four degradation modes weaken these representations, low-capacity linear probes are fit on a patch-disjoint teach split via closed-form ridge regression, and the probes' ability to recover the teacher's soft correspondence scores, inlierness, and pose on the meta split is used as an additional loss. At inference all probes and losses are removed, so the pipeline is unchanged. Experiments on 7-Scenes and RGB-D Scenes V2 report improvements over 2D3D-MATR and several older baselines, with a claimed state-of-the-art result.","tokens_in":16566,"tokens_out":6419,"duration_ms":54130,"significance":"If the reported results hold, the contribution is a novel and inexpensive regularization scheme: it requires no architectural change or inference overhead, and the idea of using weak, degraded probes as training-time constraints is original in this domain. The paper also delivers consistent internal ablations showing that each loss component and each degradation mode contributes positively, and the patch-disjoint split is a sensible precaution against trivial memorization. The main open risk is that the primary teachability signal (the correspondence loss) is defined relative to the teacher's own soft labels, so the current experiments do not yet establish that the gains come from geometric robustness rather than from a generic linear-decodability regularizer; control experiments with ground-truth-anchored probe targets are needed to close this gap. The absence of error bars and the narrow baseline set also temper the 'state-of-the-art' claim.","major_comments":[{"comment":"The correspondence teachability loss measures how well a linear probe recovers the teacher's own softened logits y_j = sigma(s_j/tau). This makes the core criterion self-referential: a representation can become easier for the probe to imitate without becoming more geometrically correct. The ablation in Table 4 shows that adding only L_corr raises RR from 75.8 to 79.6 (nearly half of the total gain to 85.2), so a large part of the reported improvement is driven by this self-referential term. To support the claim that teachability reflects robust, geometry-preserving representations, the authors should add a control experiment in which the probe targets are replaced by ground-truth correspondence labels or by the geometric inlierness q_j of Eq. (16), and compare the resulting IR/FMR/RR with the teacher-target version. If the teacher-target probe only acts as a regularizer, the current interpretation and the term 'teachability' should be recalibrated.","section":"Sec. 3.3.4 / Eq. (11), (15)"},{"comment":"The paper stops gradients through the closed-form ridge solution W* and propagates only through the representations and B_m. This means the teach split is not directly trained through the probe-fitting step; the training signal on the meta split uses a probe whose weights are treated as fixed. This is a non-trivial design choice with no ablations or analysis. The authors should either compare full differentiation through the ridge solution against the stop-gradient variant, or provide a clear justification for why the stop-gradient variant is preferred, and report any effect on the final metrics. Without this, it is unclear whether the closed-form adaptation is actually load-bearing or whether a fixed random projection would give similar results.","section":"Sec. 3.3.5 / Eqs. (12)-(13)"},{"comment":"The 'state-of-the-art' claim is supported by comparison with only five baselines, the most recent of which (Diff2I2P) is outperformed by margins of 2.2 and 2.6 percentage points in RR, with no error bars or significance tests. Run-to-run variance of this magnitude is often substantial in 2D-3D matching. The authors should report means and standard deviations over at least three seeds for the main comparisons, and include at least one recent competitive detection-free method (e.g., CorrI2P or VP2P-Match) to substantiate the state-of-the-art claim.","section":"Tables 1-2 and abstract"},{"comment":"The contribution statement claims TeaMatch 'consistently improves strong detection-free baselines', and the abstract states that it can be 'seamlessly integrated into existing coarse-to-fine matching pipelines'. However, all experiments are run on a single teacher (the 2D3D-MATR architecture; see Table 3, where the teacher baseline matches the MATR row of Table 1). No experiment applies TeaMatch to Diff2I2P or another pipeline, so the generality claim is not demonstrated. Either add experiments on a second teacher or revise the claims to be specific to the 2D3D-MATR teacher.","section":"Introduction and Sec. 4.1"}],"minor_comments":[{"comment":"The citation 'PnP-RANSAC [14?]' contains a stray question mark; reference [14] is RANSAC (Fischler and Bolles), and the '?' suggests a broken cross-reference. Please fix the citation and also cite EPnP properly where the solver is described.","section":"Sec. 3.1 and 3.5"},{"comment":"Several symbols in the figure (e.g., '?', '??', '??') are rendered as placeholder glyphs, making the architecture diagram hard to read. Provide a clean version with all notations matching the text.","section":"Figure 2"},{"comment":"The RR row for Diff2I2P appears as '99.095.6 74.098.986.8...' without spacing; reformat the table so that all numbers are clearly separated.","section":"Table 1"},{"comment":"The paper does not report the values of the hyperparameters introduced in the method (rho, tau, delta, lambda_r, lambda_t, beta_1-3, lambda_f, lambda_p). Since TeaMatch has a large number of added coefficients, please include the chosen values and a sensitivity analysis for at least the most important ones (e.g., rho and lambda_t).","section":"Sec. 4.4"},{"comment":"Many reference entries only provide a Semantic Scholar API URL instead of the original venue and DOI; please reformat the bibliography to the ACM style.","section":"References"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real, modest contribution. TeaMatch wraps a detection-free 2D-3D matcher with a training-time regularizer built from weak probes, structured degradation, and a patch-disjoint teach/meta split. The package is new relative to the probing/distillation literature, and the payoff is real: consistent gains over 2D3D-MATR and Diff2I2P on 7-Scenes and RGB-D Scenes V2, plus evidence of graceful degradation under occlusion, at zero inference cost. That is worth something.\n\nThe math is coherent. Closed-form ridge probes are fit on a teach split, evaluated on a disjoint meta split, and gradients are stopped through the probe weights; the three losses (correspondence, inlierness, pose) are sensible. The occlusion-ratio experiment is the most direct evidence that the method improves robustness, and I buy it.\n\nSoft spots, in order of importance. First, the correspondence-level loss is self-referential: the probe target is the teacher's own softened logit, so low loss means linearly predictable in the teacher's current score geometry, not necessarily geometrically correct. The inlierness and pose losses inject ground-truth geometry and partially fix this, and the ablation in Table 4 shows those terms carry a large share of the RR gain. So the narrow stress-test claim is too strong if it predicts the method is only linear decodability; the final system is anchored by GT geometry. But the paper should include an ablation where the probe is trained against GT inlier labels instead of teacher logits, to show what teachability itself buys. Second, 'state-of-the-art' is not supported: only five baselines, most old, no error bars or significance tests. Third, reproducibility is below bar: no code, and hyperparameters like rho, tau, delta, lambda_r and the loss weights are not reported. These are fixable.\n\nWho it's for: people working on detection-free image-to-point-cloud registration, and anyone interested in training-only representation regularizers. It is not a breakthrough, but it is a clean idea worth engaging with. I would send it to review, with a request for the GT-anchored probe control, error bars, and hyperparameters. If those are added, this is a solid MM paper.","headline":"A clean training-only regularizer that improves 2D-3D matching by a few percent; the core teachability signal is partly self-referential, but the geometry-anchored losses and the occlusion experiments largely back the robustness claim.","tokens_in":17073,"tokens_out":2955,"would_cite":true,"duration_ms":28900,"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":"TeaMatch claims that 2D-3D correspondences improve when the matcher's pair representations must remain recoverable by degraded weak learners on unseen patches, and reports consistent gains in inlier ratio and pose recall without changing…","keywords":["2D-3D matching","cross-modal representation learning","teachability","weak probes","teacher-student learning","visual localization","detection-free matching","pose estimation"],"falsifier":"Train the same TeaMatch pipeline but replace the probe targets $y_j = \\sigma(s_j/\\tau)$ with ground-truth geometric inlier labels $q_j = \\exp(-r_j^2/\\delta^2)$ computed from the true pose, keeping the patch-disjoint split, degradation modes, and loss weighting unchanged. If that geometrically supervised variant does not equal or beat TeaMatch's inlier ratio and registration recall, the paper's claim that teacher-score teachability is the driving signal is not supported.","tokens_in":16069,"feed_emoji":"🧭","tokens_out":11345,"duration_ms":87671,"temperature":0.7,"pith_summary":"TeaMatch claims that the quality of image-to-point-cloud representations for 2D-3D matching can be improved by demanding that they remain recoverable under degraded conditions. The paper defines this property as teachability and enforces it with low-capacity probes that imitate the teacher's softened matching scores on one set of coarse patches and are evaluated on a disjoint set. All probes and auxiliary losses exist only during training, so the framework can be layered onto a standard detection-free coarse-to-fine matcher with no change at inference time. On the 7-Scenes and RGB-D Scenes V2 benchmarks, the authors report consistent improvements in inlier ratio, feature matching recall, and registration recall, including under severe point-cloud occlusion. If the claim is right, cross-modal matching robustness can be bought entirely through training-time representation learning rather than architectural changes.","feed_headline":"Weak learners boost 2D-3D pose recall to 85.2%","feed_subtitle":"Training-time weak probes make images and point clouds match more reliably, with zero added test-time cost.","key_machinery":"The load-bearing object is the pair-level multimodal representation $z_j = \\phi(f^I_j, f^P_j, c^I_j, c^P_j, \\hat{u}_j, \\hat{x}_j)$, built from local image and point features, coarse contextual features, and normalized spatial coordinates. TeaMatch degrades $z_j$ through image weakening, geometry weakening, context removal, or a low-rank bottleneck, then fits a weak ridge-regression probe in closed form, $W^*_m = ((H^{(m)}_{\\mathrm{teach}})^\\top H^{(m)}_{\\mathrm{teach}} + \\lambda_r I)^{-1} (H^{(m)}_{\\mathrm{teach}})^\\top Y_{\\mathrm{teach}}$, on a teach split and evaluates it on a patch-disjoint meta split. The meta-split probe predictions feed three losses, $L^{(m)}_{\\mathrm{corr}}$, $L^{(m)}_{\\mathrm{inlier}}$, and $L^{(m)}_{\\mathrm{pose}}$, which are combined as $L_{\\mathrm{teach}} = \\frac{1}{|\\mathcal{S}|}\\sum_m (\\beta_1 L^{(m)}_{\\mathrm{corr}} + \\beta_2 L^{(m)}_{\\mathrm{inlier}} + \\beta_3 L^{(m)}_{\\mathrm{pose}})$ and added to the task loss as $L = L_{\\mathrm{task}} + \\lambda_t L_{\\mathrm{teach}}$. Gradients are stopped through the closed-form probe solution and propagated only through the teacher representations, so the probes act as pure regularizers.","core_discovery":"TeaMatch's central claim is that a pair-level representation is good for 2D-3D matching only if a deliberately weak learner can recover the teacher's matching preference from it after structured degradation, and only on patch regions the learner did not imitate. The paper turns this criterion into a training objective: degrade candidate pair representations, fit ridge-regression probes in closed form on a teach split of coarse patch pairs, evaluate them on a disjoint meta split, and supervise the teacher with three losses—correspondence-score recovery, soft inlier ratio against ground-truth geometry, and differentiable PnP pose alignment. The authors report that this teachability regularization lifts mean inlier ratio from 50.1% to 54.6% and registration recall from 75.8% to 85.2% on 7-Scenes over the teacher baseline, and improves all metrics on RGB-D Scenes V2, with the auxiliary machinery removed at inference.","pith_inferences":["Because the training recipe is agnostic to the modality pair, the same teachability objective could be transferred to other correspondence problems, such as image-image or point-cloud-point-cloud matching, whenever candidate pair representations and soft matching targets exist; the paper does not test these settings.","The correspondence-level loss only needs the teacher's own soft scores, not ground-truth pose, so a semi-supervised variant could apply teachability to large unlabeled image-point collections and reserve the inlierness and pose losses for labeled subsets; the paper does not propose this.","A natural stress test would be to evaluate TeaMatch-trained representations under degradation types absent from the four training modes; if the gains are a genuine property of the representation, they should transfer to novel corruptions, but the paper only reports the trained modes.","Making the degradation family adaptive—selecting, per mini-batch, the mode that most hurts probe recovery—could push the representation harder than the fixed set of four modes studied here."],"forward_implications":["Because all teachability components are dropped at inference, the framework can be added to any existing detection-free coarse-to-fine matcher with no added latency, memory, or model size.","On 7-Scenes the mean inlier ratio rises from 50.1% to 54.6% and registration recall from 75.8% to 85.2% relative to the teacher baseline.","On RGB-D Scenes V2 the method reaches 63.1% registration recall and 78.3% feature matching recall, showing the benefit carries across unseen scenes.","Under point-cloud occlusion up to 60%, the inlier ratio remains at 44.2% while the strongest prior baseline drops to 23.5%, indicating more graceful degradation.","Ablations show the gains depend on all three teachability losses and on the patch-disjoint teach/meta split; removing them or splitting at random lowers performance."],"supporting_citations":[{"why":"It supplies the detection-free coarse-to-fine matcher used as the teacher backbone and as the baseline against which TeaMatch's main gains on 7-Scenes are measured.","marker":"[32]"},{"why":"It supplies a leading prior detection-free matcher whose results on both benchmarks are the main comparison points for TeaMatch's improvements and robustness curves.","marker":"[36]"},{"why":"It supplies the 7-Scenes benchmark and its official train/validation/test split, the primary evaluation set for the main tables and ablations.","marker":"[16]"},{"why":"It supplies the RGB-D Scenes V2 benchmark used to test whether TeaMatch's gains generalize across unseen scenes.","marker":"[26]"},{"why":"It supplies the linear-probe methodology this work adapts into training-time weak learners that assess representation recoverability.","marker":"[1]"},{"why":"It supplies the knowledge-distillation background that motivates learning from a teacher, which TeaMatch reframes as a recoverability regularizer rather than a deployed student.","marker":"[21]"},{"why":"It supplies the ResNet-FPN image encoder used in the shared backbone, fixing the architecture that the teachability objective regularizes.","marker":"[19]"},{"why":"It supplies the KPConv-FPN point-cloud encoder used in the shared backbone, fixing the architecture that the teachability objective regularizes.","marker":"[50]"}],"fun_headline_variants":["Teachability lifts 2D-3D registration recall to 85.2%","Weak-student probes sharpen 2D-3D matching at zero cost","Training with weak learners boosts 2D-3D pose accuracy by 9.4%","Meta-split teaching: robust 2D-3D matching without test-time overhead","TeaMatch: teachable representations for reliable 2D-3D matching"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the assumption that 'a weak learner's ability to recover the teacher's own softened matching scores on degraded, unseen image-point patches' is a faithful measure of whether those representations will support geometrically correct correspondences and poses.","fun_headline_variants_meta":{"raw":{"variants":["Teachability lifts 2D-3D registration recall to 85.2%","Weak-student probes sharpen 2D-3D matching at zero cost","Training with weak learners boosts 2D-3D pose accuracy by 9.4%","Meta-split teaching: robust 2D-3D matching without test-time overhead","TeaMatch: teachable representations for reliable 2D-3D matching"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1373,"prompt_tokens":939,"completion_tokens":434,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":324}},"tokens_in":555,"tokens_out":434,"duration_ms":4372,"temperature":1.0,"reasoning_tokens":324,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:30:15.556274+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same TeaMatch pipeline but replace the probe targets $y_j = \\sigma(s_j/\\tau)$ with ground-truth geometric inlier labels $q_j = \\exp(-r_j^2/\\delta^2)$ computed from the true pose, keeping the patch-disjoint split, degradation modes, and loss weighting unchanged. If that geometrically supervised variant does not equal or beat TeaMatch's inlier ratio and registration recall, the paper's claim that teacher-score teachability is the driving signal is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the detection-free coarse-to-fine matcher used as the teacher backbone and as the baseline against which TeaMatch's main gains on 7-Scenes are measured."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies a leading prior detection-free matcher whose results on both benchmarks are the main comparison points for TeaMatch's improvements and robustness curves."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the 7-Scenes benchmark and its official train/validation/test split, the primary evaluation set for the main tables and ablations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the RGB-D Scenes V2 benchmark used to test whether TeaMatch's gains generalize across unseen scenes."},{"cited_title":"Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, François Goulette, and Leonidas J","cited_arxiv_id":null,"evidence_quote":"It supplies the KPConv-FPN point-cloud encoder used in the shared backbone, fixing the architecture that the teachability objective regularizes."}],"review_version":1}