{"id":"7fee71f0-3abc-4479-b223-6ab0394a843c","arxiv_id":"2504.19643","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"BARIS-ERA, combining a gated multi-scale decoder, a parameter-efficient environmental adapter, and a boundary-aware loss, reports state-of-the-art underwater instance segmentation on UIIS and USIS10K.","lead":"A new decoder and a lightweight adapter help computers outline fish and other objects in murky underwater photos, improving accuracy by three to four points on public benchmarks. The approach adapts a large pretrained model to underwater conditions while training under five percent of the parameters, a practical trade-off for underwater robotics.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BACE loss's range-null-space derivation is invalid as implemented: max-pooling is nonlinear and has no pseudo-inverse, so the theoretical support for this named contribution collapses unless a corrected test with a linear operator preserves the gain.","rationale":"The reader and I converge on the same load-bearing soft spot. The paper's headline SOTA numbers could in principle survive a flawed loss derivation, but BACE Loss is one of the three named contributions and Section 3.3.1 is the only theoretical argument for it. The implementation uses max-pooling, which is nonlinear and hence has no pseudo-inverse, so the range-null-space decomposition presented in Eqs. (13)-(15) does not describe the implemented loss. This is an internal inconsistency rather than a disagreement with an external convention. The proposed average-pooling test cleanly separates the empirical from the theoretical: with average pooling the same formulas become mathematically valid, so if the mAP gain vanishes, the claim that BACE improves segmentation via range-null-space refinement is unsupported. I do not recommend rejection because the ablations are internally consistent and the decoder and ERA components are plausible engineering contributions, but the loss's theoretical framing must be corrected or re-validated. Secondary concerns (no public code or weights, no variance estimates, hyperparameters γ, N, and block count selected on the same validation split) reinforce the conditional verdict and should be addressed for the SOTA claim to be fully reproducible.","tokens_in":16683,"tokens_out":11187,"duration_ms":118904,"concrete_test":"Re-run the Table 6 and Table 4 BACE ablations with A implemented as average pooling over the same k×k window (keeping A^T as nearest-neighbor upsampling). For non-overlapping average pooling, A is linear and A^T A is the orthogonal projection onto the space of piecewise-constant masks, making Eqs. (13)-(15) valid. If the +1.1 mAP BACE gain (and the contribution to the full model) disappears or reverses, the reported improvement depends on the nonlinear max-pooling operator, not on range-null-space decomposition; if the gain persists, the loss is empirically useful but requires a corrected theoretical justification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3.1 rests BACE Loss on Eqs. (13)-(15): a linear operator A with pseudo-inverse A† satisfying AA†A=A, and the decomposition x = A†Ax + (I-A†A)x. The implementation in Section 3.3.2 and Appendix 6.3 sets A = MaxPooling and A^T = nearest-neighbor Upsample. Max pooling is not a linear map, so it has no matrix representation, no pseudo-inverse, and Eqs. (13)-(15) do not apply; the 'range-space' and 'null-space' language is therefore not justified. The composed operator P = A^T A is idempotent for non-overlapping pooling, but it is a nonlinear morphological dilation (for binary masks) rather than an orthogonal projection. Eq. (16) also appears to reverse the BCE argument order relative to the PyTorch code, which calls BCEWithLogits(refined_pred, label). Because Table 4 shows BACE alone improves mAP by 1.1 (28.2 to 29.3) and the full model includes λ·L_BACE, the flawed derivation is load-bearing for the paper's stated novelty, even though the empirical gain could in principle survive a corrected explanation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BARIS, a Mask R-CNN extension for underwater instance segmentation, with three main contributions: a BARIS decoder built from a Multi-Stage Gated Refinement Network and Depthwise Separable Upsampling; an Environmental Robust Adapter (ERA) that adds learnable environmental embeddings and multi-scale feature extraction to a frozen backbone; and a Boundary-Aware Cross-Entropy (BACE) loss that is claimed to refine mask boundaries via range-null space decomposition. The authors report state-of-the-art results on the UIIS dataset (31.6 mAP with Swin-B, 32.3 with ConvNeXt V2-B) and on USIS10K (47.3 mAP), with reductions of over 90% in trainable backbone parameters relative to full fine-tuning. The paper includes extensive ablations over the decoder, adapter, loss, number of refinement blocks, projection ratio, and number of environmental embeddings, as well as FPS comparisons.","tokens_in":16997,"tokens_out":4939,"duration_ms":53682,"significance":"If the empirical claims hold, the work is a useful contribution to parameter-efficient underwater instance segmentation: it demonstrates that a frozen pretrained backbone with a lightweight adapter and a refinement decoder can outperform full fine-tuning and match or exceed a much larger ViT-H-based model. The strengths are the external benchmarks, the per-component ablations in Tables 4-8, the parameter-efficiency comparisons in Tables 3 and 10, and the explicit reporting of FPS in Table 9. However, the theoretical framing of the BACE loss as a range-null space projection is not supported by the implementation, and the final configuration is selected on the same UIIS evaluation set. These issues do not necessarily overturn the empirical results, but they currently block acceptance because the named novelty of the BACE loss rests on an invalid derivation and the headline SOTA numbers may be partially an artifact of benchmark-specific tuning.","major_comments":[{"comment":"The range-null-space derivation does not apply to the implemented loss. Equation (13) defines a pseudo-inverse A† satisfying AA†A=A, but Equation (15) replaces A† with A^T; A^T A is the orthogonal projector onto the range of A only when the rows of A are orthonormal, which is not assumed or stated. More seriously, Section 3.3.2 and Appendix 6.3 set A to max-pooling and A^T to nearest-neighbor upsampling; max-pooling is nonlinear and has no matrix representation and no pseudo-inverse, so the decomposed projections in Equations (14) and (15) are undefined. Since Table 4 attributes +1.1 mAP to BACE alone, this issue is load-bearing for the paper's stated novelty. Please either reformulate BACE as a heuristic boundary-refinement loss without the projection claim, or provide an implementation with a genuinely linear operator (e.g., average pooling or blur) using the correct pseudo-inverse A† and show that the gain persists.","section":"Section 3.3.1, Eq. (13)-(15) and Appendix 6.3"},{"comment":"The BACE formula reverses the BCE argument order relative to the code in Appendix 6.3. Equation (16) writes BCE(Mgt, Γ(Mθ,Mgt)), but the PyTorch-like code calls BinaryCrossEntropyWithLogits(refined_pred, label), which means the refined prediction should be the first argument and the ground truth the second, i.e., BCE(Γ(Mθ,Mgt), Mgt). Please correct the formula and confirm that the ablation in Table 4 corresponds to the corrected ordering.","section":"Section 3.3.3, Eq. (16)"},{"comment":"The final configuration is selected on the same evaluation benchmark used for the SOTA comparison. Tables 7, 8, and 11 select the number of refine blocks, the projection ratio γ, and the number of environmental embeddings by comparing UIIS mAP, and the best settings are then used for the headline results in Table 1 on the same UIIS set. This makes the reported gains partly a result of benchmark-specific tuning rather than a pre-registered evaluation. Please describe the tuning protocol explicitly, use a held-out validation split for model selection, or otherwise demonstrate that the improvement over baselines is not driven by selection on the evaluation set.","section":"Section 4.4, Tables 7, 8, 11"}],"minor_comments":[{"comment":"The supplementary training setup refers to 'the RefineMask module' when describing the proposed architecture; this should be 'the BARIS-Decoder module' to avoid confusion with the RefineMask baseline.","section":"Section 6.1, Appendix"},{"comment":"The text says the number of ERA parameters was adjusted so that its parameter count 'closely matches' that of MONA, but Table 3 reports MONA at 3.67M and ERA at 4.25M; please clarify whether the comparison is parameter-matched and, if so, why the counts differ.","section":"Table 3 and Section 4.3"},{"comment":"The notation is inconsistent: Xn is defined with DSConv 3×3, while the MSGAttention formula uses a separate W = DSConv(X'1) that is never connected to the attention computation; please clarify the role of W in the gated attention and align the symbols with Figure 3.","section":"Section 3.1.1, Eq. (3)-(4)"},{"comment":"BARIS-ERA runs at 4.866 FPS versus 8.325 FPS for Mask R-CNN with Swin-B, a 42% slowdown; the text calls this 'competitive', which should be qualified given the substantial speed cost.","section":"Section 6.4, Table 9"},{"comment":"No variance or multi-seed statistics are reported for the mAP values; given that several advertised gains are 1-2 mAP, please report standard deviations or at least note that results are from a single run.","section":"Section 4.1 and Tables 1-2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, the short version: this is a solid engineering paper with plausible benchmark numbers, but its headline novelty—the BACE loss—is supported by a derivation that does not match the implementation. As written, the math only works for linear A, and the code uses max-pooling. That's an internal contradiction, not a nitpick.\n\nWhat's actually new: the specific assembly of a RefineMask-style multi-scale decoder, a Mona-Tuning-like adapter with learned environment embeddings, and a boundary-aware loss. The gains on UIIS (+3.4 mAP over Mask R-CNN with Swin-B) and USIS10K (+4.2 over USIS-SAM) are plausible and the ablations are thorough. The parameter-efficiency story is genuine: ERA uses ~4.7% of backbone trainable parameters while beating full fine-tuning. The comparisons to existing fine-tuning methods and the t-SNE analysis are useful. So there is real engineering here.\n\nThe big soft spot is the BACE loss. Section 3.3.1 invokes range-null space decomposition for a linear operator A with pseudo-inverse, and Eq. (15) requires A^T A to be an orthogonal projection. The implementation in Section 3.3.2/Appendix 6.3 sets A to max-pooling, which is nonlinear and has no pseudo-inverse or projection interpretation. So the theoretical support for the named component collapses. That matters because Table 4 credits BACE alone with +1.1 mAP. The loss might still work empirically—the composition of max-pooling and nearest-neighbor upsampling can act like a smoothing operator—but the paper needs a corrected explanation or an ablation with an actual linear A.\n\nSecondary issues: the number of environment embeddings (Table 11) and projection ratio (Table 8) are tuned on the validation set, which inflates the reported numbers; there are no variance estimates or significance tests; and no code/weights are provided. Also, Eq. (16) has the BCE argument order reversed relative to the code—minor, but symptomatic of sloppiness.\n\nIs it worth a referee? Yes, if the authors fix the BACE theory and release the implementation. The empirical claim is strong enough to warrant a proper review, but the current version's central theoretical argument is wrong. I'd send it back for major revision rather than desk-reject, and I'd ask a reviewer to check the BACE ablation carefully.","headline":"Plausible mAP gains but a load-bearing mathematical flaw in the BACE loss derivation; referee only if the authors fix the theory and release code.","tokens_in":17489,"tokens_out":3356,"would_cite":false,"duration_ms":32979,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that underwater instance segmentation can be substantially improved by combining a boundary-focused decoder with a lightweight adapter that learns environmental degradation priors, while freezing most of the backbone.","keywords":["underwater instance segmentation","boundary-aware refinement","adapter tuning","range-null space decomposition","multi-scale gated attention","degradation priors","Mask R-CNN","BACE loss"],"falsifier":"Test the identity $AA^\\dagger A = A$ on the actual implementation: apply max pooling followed by nearest-neighbor upsampling to a random mask; if the composition does not reproduce the pooled-and-restored layout (it generally will not), the range-null space justification does not hold for the code. Then run the same BACE loss with average pooling replacing max pooling; if the mAP gain persists, the linear-theory interpretation is confirmed as contributing, and if it disappears, the nonlinearity itself is what the loss was exploiting.","tokens_in":16467,"feed_emoji":"🌊","tokens_out":5706,"duration_ms":48382,"temperature":0.7,"pith_summary":"The paper proposes BARIS, a framework for underwater instance segmentation built from three components: a boundary-aware decoder (BARIS-Decoder), an adapter that learns environmental degradation priors (ERA), and a boundary-aware cross-entropy loss (BACE Loss). It claims that together these components achieve state-of-the-art results on two underwater benchmarks, surpassing Mask R-CNN by 3.4 mAP with a Swin-B backbone and 3.8 mAP with ConvNeXt V2 on UIIS, and outperforming the heavier USIS-SAM (ViT-H) by 4.2 mAP on USIS10K. The paper further claims that ERA cuts trainable backbone parameters by over 90% while still beating full fine-tuning, suggesting that explicit boundary refinement and learned degradation priors, rather than larger models, drive robust underwater segmentation.","feed_headline":"Underwater mask AP up 3.4–3.8 with 90% fewer trainable params","feed_subtitle":"A gated refinement decoder plus degradation-prior adapter beats Mask R-CNN, WaterMask, and a ViT-H SAM baseline.","key_machinery":"The load-bearing machinery has three parts. The Multi-Stage Gated Refinement Network (MSGRN) fuses features from all four backbone stages through multi-scale gated attention and residual connections. The Depthwise Separable Upsample (DSU) replaces bilinear upsampling with multi-scale depthwise convolutions and pixel shuffle. The Environmental Robust Adapter (ERA) is inserted at the end of each frozen transformer or convolutional block; it uses multi-scale feature extraction, channel attention, and a set of learnable environmental embeddings $E \\in \\mathbb{R}^{N \\times C}$ that softly gate each pixel's features, followed by a zero-initialized up-projection. The Boundary-Aware Cross-Entropy (BACE) Loss refines the predicted mask with $\\Gamma(M_\\theta, M_{gt}) = A^T A M_{gt} + (I - A^T A) M_\\theta$, which the paper interprets as a range-null space decomposition.","core_discovery":"On its own terms, the paper establishes that BARIS-ERA achieves state-of-the-art underwater instance segmentation: it surpasses Mask R-CNN by 3.4 mAP with a Swin-B backbone and 3.8 mAP with ConvNeXt V2-B on UIIS, and it outperforms USIS-SAM (ViT-H) by 4.2 mAP on USIS10K, while training only about 4% of the backbone parameters. The ablations attribute 1.8 mAP of the gain to the BARIS-Decoder, 2.0 mAP to ERA, and 1.1 mAP to BACE Loss, with the full model reaching 31.6 mAP on UIIS. The paper frames this as evidence that explicit boundary refinement plus learned environmental degradation priors, rather than larger backbones, are what make underwater segmentation robust.","pith_inferences":["Because the BACE loss benefit is attributed to a decomposition that the implemented operators do not satisfy, a cleaner test would be to replace max pooling with a genuinely linear operator such as average pooling or a fixed blur and check whether the loss still helps; this would isolate whether the mechanism is the linear algebra or the nonlinear sharpening.","The learnable environmental embeddings act as per-pixel soft gating over a small set of degradation prototypes; a natural extension is to probe whether the same embeddings transfer to other degraded domains such as fog, haze, or low light without retraining.","BARIS-ERA is only evaluated inside Mask R-CNN; combining the decoder and adapter with query-based segmenters such as Mask2Former or SAM-based pipelines could reveal whether the boundary refinement generalizes across segmentation paradigms."],"forward_implications":["Adapter-based tuning with environmental priors can beat full fine-tuning for underwater instance segmentation: ERA reaches 29.9 mAP against 28.2 for full fine-tuning while training only 4.67% of backbone parameters.","BACE Loss adds 1.1 mAP over plain cross-entropy on UIIS and outperforms two prior boundary-aware losses, PIDNet-style boundary awareness and Active Boundary Loss.","The gains are consistent across architectures: the best mAP is 31.6 with Swin-B using projection ratio $\\gamma=2$, and 32.3 with ConvNeXt V2-B using $\\gamma=4$.","BARIS-ERA keeps inference practical: it runs at 4.866 FPS with 114.44 M parameters, versus 2.750 FPS and 698.12 M parameters for USIS-SAM.","Design choices have measured sweet spots: three refine blocks and 16 environmental embeddings maximize mAP on UIIS."],"supporting_citations":[{"why":"Mask R-CNN baseline; the base detector that BARIS-ERA turns into a state-of-the-art underwater segmenter, providing the 28.2 mAP reference point.","marker":"[15]"},{"why":"WaterMask; the strongest previous underwater instance segmentation method and the source of the UIIS dataset, beaten by 1.5 mAP.","marker":"[24]"},{"why":"USIS-SAM; the adapter-based underwater method with a ViT-H backbone and the source of the USIS10K dataset, beaten by 4.2 mAP.","marker":"[25]"},{"why":"RefineMask; the multi-scale refinement approach that BARIS-Decoder extends and outperforms in the refinement ablation.","marker":"[48]"},{"why":"MONA; the adapter-tuning baseline that ERA is inspired by and surpasses at a matched parameter budget.","marker":"[45]"},{"why":"SegAdapter; supplies the High-Order Spatial Attention that MSGAttention adapts for multi-scale gating.","marker":"[34]"},{"why":"Null-space and range-null space methods from inverse problems; the mathematical basis the BACE loss claims.","marker":"[41, 42]"}],"fun_headline_variants":["BARIS-ERA: underwater seg +3.4-3.8 mAP, 90% fewer tuned params","Underwater seg: 3.4-3.8 mAP gain, 90% param reduction via ERA adapter","BARIS-ERA beats Mask R-CNN by 3.4-3.8 mAP underwater, 90% fewer trained weights","Underwater instance segmentation: 3.8 mAP gain over Mask R-CNN with 90% less tuning","Outperforms ViT-H SAM by 4.2 mAP on USIS10K with 90% fewer tuned params"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theoretical derivation of the BACE loss assumes a linear operator $A$ with a pseudo-inverse satisfying $AA^\\dagger A = A$, but the implementation uses max pooling as $A$ and nearest-neighbor upsampling as $A^T$, and max pooling is nonlinear and has no such pseudo-inverse.","fun_headline_variants_meta":{"raw":{"variants":["BARIS-ERA: underwater seg +3.4-3.8 mAP, 90% fewer tuned params","Underwater seg: 3.4-3.8 mAP gain, 90% param reduction via ERA adapter","BARIS-ERA beats Mask R-CNN by 3.4-3.8 mAP underwater, 90% fewer trained weights","Underwater instance segmentation: 3.8 mAP gain over Mask R-CNN with 90% less tuning","Outperforms ViT-H SAM by 4.2 mAP on USIS10K with 90% fewer tuned params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001933,"raw_usage":{"total_tokens":7539,"prompt_tokens":895,"completion_tokens":6644,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":6493}},"tokens_in":511,"tokens_out":6644,"duration_ms":47723,"temperature":1.0,"reasoning_tokens":6493,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:46:40.895345+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Test the identity $AA^\\dagger A = A$ on the actual implementation: apply max pooling followed by nearest-neighbor upsampling to a random mask; if the composition does not reproduce the pooled-and-restored layout (it generally will not), the range-null space justification does not hold for the code. Then run the same BACE loss with average pooling replacing max pooling; if the mAP gain persists, the linear-theory interpretation is confirmed as contributing, and if it disappears, the nonlinearity itself is what the loss was exploiting.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Mask R-CNN baseline; the base detector that BARIS-ERA turns into a state-of-the-art underwater segmenter, providing the 28.2 mAP reference point."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"WaterMask; the strongest previous underwater instance segmentation method and the source of the UIIS dataset, beaten by 1.5 mAP."},{"cited_title":"Lian and others","cited_arxiv_id":null,"evidence_quote":"USIS-SAM; the adapter-based underwater method with a ViT-H backbone and the source of the USIS10K dataset, beaten by 4.2 mAP."},{"cited_title":"Zhang, X","cited_arxiv_id":null,"evidence_quote":"RefineMask; the multi-scale refinement approach that BARIS-Decoder extends and outperforms in the refinement ablation."},{"cited_title":"Peng and W","cited_arxiv_id":null,"evidence_quote":"SegAdapter; supplies the High-Order Spatial Attention that MSGAttention adapts for multi-scale gating."}],"review_version":1}