{"id":"699ad5bb-5aae-42e9-bad0-36f54cfbddd8","arxiv_id":"2508.06063","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A shared network with about 2,000 decoder-specific parameters and a saliency-filtered, size-balanced training set reaches state-of-the-art accuracy on both salient and camouflaged object detection simultaneously.","lead":"This paper trains one network that performs both salient object detection and camouflaged object detection, two tasks previously thought too contradictory to share a model. Two small additions, a per-task normalization dial in the decoder and a saliency-ranked data filter, let both tasks improve together.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing control: no independent training on the SBSS-filtered subset, so the reported 'joint learning benefit' is confounded by data selection.","rationale":"The paper is internally consistent and the ablations are extensive, but the central claim—that a fully shared network with ~2K task-specific parameters lets both SOD and COD benefit from joint training—rests on a comparison that does not isolate the joint-training mechanism. The reader identified the SBSS oracle/similarity proxy as the weakest assumption; I agree that this is a genuine concern, but the more load-bearing issue is the missing control. Even granting that SENet's ranking perfectly measures image-GT quality, the reported SOD improvement over IT could still be due to the fact that the training set was replaced by a carefully selected 38% subset. The architecture mismatch is equally important: IT is plain SENet, while JoNet includes DLM parameters, so the COD improvement may not require SOD data at all. The paper's own admission in Section IV-C that performance 'mainly comes from our powerful baseline model SENet' and the tension between Fig. 2 ('small fraction' low-quality) and the 61.7% discard rate reinforce that SBSS is primarily a subsampling/size-balancing mechanism. The proposed concrete test—independent training on Dsub_S and on DC with the same architecture—would directly resolve whether the gains are joint-learning effects or selection/architecture artifacts. The verdict remains CONDITIONAL: the claim is plausible and well-ablated internally, but this missing control must be supplied before the attribution is secure. I do not move to ACCEPT or REJECT because the evidence is suggestive and the test could go either way.","tokens_in":22153,"tokens_out":5715,"duration_ms":68967,"concrete_test":"Train three models with identical settings (MAE init, 25 epochs, 384×384): (1) JoNet-SOD-only: full JoNet architecture (SENet + SOD DLM) trained only on Dsub_S (top-4040 SBSS subset); (2) JoNet-COD-only: same architecture (COD DLM) trained only on the 4040 COD pairs; (3) JoNet-Random: SCJoint with Random-K instead of Top-K, for completeness. Compare (1) against JoNet's SOD columns in Table I and (2) against JoNet's COD columns. If (1) reaches DUTS Sα ≈0.931 and PASCAL-S Sα ≈0.887 without seeing COD data, the SOD-side 'joint benefit' is a selection effect; if (2) reaches CAMO Sα ≈0.889, the COD-side gain is an architecture effect. The joint-learning claim stands only if both are clearly below JoNet on their respective tasks.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section III-C and Algorithm 1 define SBSS: rank all 10,553 DUTS pairs by S = Sα + Eφ + Fβ + (1−M) (Eq. 7) using SENet's prediction-GT similarity, then keep the top 4,040. Table I compares JoNet (JT+SBSS+SCJoint) against IT, where IT is SENet trained independently on the full 10,553 SOD images and the full COD set. This comparison changes two variables at once: (i) the SOD training set becomes the top-4,040 selected subset, and (ii) the architecture gains DLM parameters plus COD supervision. No row trains the JoNet architecture on Dsub_S alone or on DC alone. Consequently, the SOD gains (DUTS Sα .921→.931; PASCAL-S .883→.887) could be due entirely to selecting high-quality/easy samples, not to joint learning. Similarly, the COD gains (CAMO Sα .875→.889) could come from the DLM module itself rather than from SOD supervision, since IT uses plain SENet without DLMs. The reader's oracle concern is real but secondary: even if SENet's agreement is a perfect quality measure, SBSS is also a size-reduction mechanism—it discards 61.7% of DUTS to match COD's count, while Fig. 2 suggests low-quality pairs are only a small fraction. Without an independent-training control on the exact filtered subset, the paper's central claim 'both tasks benefit from joint learning' is not identified; it may be a data-curation benefit plus an architecture change.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes JoNet, a single Transformer encoder-decoder for joint salient object detection (SOD) and camouflaged object detection (COD). SCJoint inserts per-task Distribution Learning Modules (DLMs) into every decoder block; each DLM is a pair of learnable mean/variance vectors, so the two tasks share all other parameters and differ by roughly 2K parameters. To balance the larger SOD training set with the smaller COD set, SBSS ranks all 10,553 DUTS image-GT pairs by the composite saliency score S = Sα + Eφ + Fβ + (1−M) computed from a pretrained SENet, and keeps the top 4,040 pairs. The paper reports ablation studies (Tables I–V) and state-of-the-art comparisons on five SOD and four COD benchmarks (Tables VI–VII). The central claim, stated in the abstract and Section I, is that with the correct joint-learning scheme, a fully shared network can outperform independent task-specific training on both tasks.","tokens_in":22468,"tokens_out":7175,"duration_ms":81543,"significance":"If the central claim is established, this is a significant result: it contradicts the commonly held view that SOD and COD joint training confuses the network, and it does so with a very cheap mechanism (~2K task-specific parameters). The ablation design is a genuine strength: Table I decomposes independent training, naive joint training, SBSS, and SCJoint; Table II explores DLM placement; Table III compares top/bottom/random sampling; Table IV sweeps data ratios; and Table V compares three joint-learning strategies on a common backbone. The authors also release code and openly acknowledge that part of the performance comes from the strong SENet baseline. The weakness is that the headline comparison changes three variables at once, so the joint-learning benefit is not yet cleanly identified. The empirical claims are plausible but require additional controls before the central conclusion can be accepted.","major_comments":[{"comment":"The headline comparison IT vs. JT+SBSS+SCJoint changes three variables simultaneously: (i) the SOD training set (full DS vs. the top-4040 subset Dsub_S), (ii) the architecture (DLMs added), and (iii) the training regime (independent vs. joint). No row trains plain SENet independently on Dsub_S, and no row trains the JoNet architecture independently on Dsub_S or on DC alone. Therefore the SOD gains (DUTS Sα .921→.931; PASCAL-S .883→.887) could be entirely due to selecting the 4,040 high-scoring/easy pairs, and the COD gains (CAMO Sα .875→.889; COD10K Sα .865→.871) could come from the DLM module or from exposure to a cleaner SOD subset rather than from joint learning as such. Please add the missing controls: (a) SENet trained independently on Dsub_S; (b) JoNet with only the SOD DLM trained independently on DS and on Dsub_S; (c) JoNet with only the COD DLM trained independently on DC; and i","section":"Table I, Section IV-B"},{"comment":"SBSS ranks image-GT pairs using S = Sα + Eφ + Fβ + (1−M), which is the same composite of the four metrics used to report all results, and the ranking model is SENet [3], the authors' own prior work. The paper assumes that higher prediction-GT similarity implies higher pair quality, but this may instead track sample easiness or distribution overlap with the test sets. Table III's Top-K/Bottom-K/Random-K comparison is informative, but it is run only under the full joint training with SCJoint, so it does not isolate the selection effect from the architecture and joint-supervision effects. Please report the distribution of S over DS, show quantitative properties of retained vs. discarded pairs (object size, contrast, etc.), and provide the independent-training control on Dsub_S requested above. This would address the concern that the reported gains are a selection artifact rather than a join","section":"Section III-C, Eq. (7), Algorithm 1"},{"comment":"The SOTA comparisons are not same-data comparisons: JoNet for SOD is trained with additional COD supervision (4,040 camouflaged images), and JoNet for COD is trained with a filtered 4,040-image SOD subset, whereas the listed task-specific specialists are trained only on their own task's training set. This does not invalidate the results, but the claim of 'outperforming task-specific specialist models' should be qualified. A fairer comparison would include a specialist SENet variant trained on the same additional data or on the same filtered subset, at least as a reference point. The current Table I baselines are the right starting point, but the missing controls from the first comment are needed before the SOTA framing is fully supported.","section":"Tables VI–VII, Section IV-C"}],"minor_comments":[{"comment":"The notation paragraph says '(µs, σs), (µc, σc) to represent the two sets of task-specific parameters for COD and SOD respectively', which appears to be swapped: Eq. (1) uses (µs, σs) for SOD and Eq. (2) uses (µc, σc) for COD. Please correct.","section":"Section III-B"},{"comment":"The composite score S = Sα + Eφ + Fβ + (1−M) uses Fβ, but the paper otherwise distinguishes maximum F-measure for SOD and weighted F-measure for COD. Since SBSS is applied to the SOD training set, please state explicitly which Fβ variant is used in Eq. (7) and justify the equal weighting of the four terms.","section":"Eq. (7), Section IV-A"},{"comment":"The pseudocode says 'Use pretrained SOD network f' but Section III-C specifies SENet. Please make the identity of f explicit in the algorithm caption, especially because the choice of oracle is central to SBSS.","section":"Algorithm 1"},{"comment":"The three-task generalization experiment uses randomly selected 4,040 images per task; for the SD task, SBU contains 4,089 training images, so the selection is not random over the full set and the effect of this subsampling is not controlled. Also, the OOD claim is based solely on GPT-4-generated qualitative examples; please temper the wording or provide a quantitative OOD evaluation.","section":"Section IV-F, Fig. 9"},{"comment":"In Table VI, the row for JoNet has a dash in the Venue column and no reference; please add the submission venue or mark it as 'Ours' consistently with the other tables.","section":"Table VI"},{"comment":"The observation that inserting DLM in the encoder 'brings devastating damage' to SOD while preserving COD performance is interesting but unexplained. A sentence discussing a possible mechanism (e.g., interaction with pretrained MAE features or gradient imbalance) would strengthen the paper.","section":"Section IV-B, DLM location"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically solid and the ablation is unusually thorough, but the central claim is currently under-identified because the main comparison changes data selection, architecture, and training regime at once. The missing independent-training control on the SBSS-filtered subset is the load-bearing issue. A second concern is the use of the authors' own SENet as the ranking oracle with a score composed of the exact evaluation metrics; this is not disqualifying, but it raises a selection-bias risk that should be addressed with an external oracle or additional analysis. I recommend major revision rather than rejection because the missing experiments are well within the scope of the paper and the proposed fixes are straightforward."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth knowing: this paper shows that a single ViT with about 2K task-specific affine parameters (the DLM) plus a saliency-based data filter (SBSS) reaches SOTA on both SOD and COD, beating task-specific specialists. That is a real system-level result, and the ablation suite is unusually thorough: IT vs JT vs each component, DLM placement, sampling direction, data-ratio sweeps, and a fair comparison of four joint-learning strategies on one backbone. They also show the trick generalizes to shadow detection. Credit where due: the paper is honest that the strong SENet baseline drives much of the performance, and it includes failure cases.\n\nThe soft spot is attribution. Table I compares JoNet (joint training + DLM + SBSS-filtered SOD subset) against IT (plain SENet, full data, independent training). That changes three variables at once, so the claim that \"both tasks benefit from joint learning\" is not actually identified. The missing controls are straightforward: train the JoNet architecture independently on the full SOD set, on the SBSS-filtered SOD subset alone, and on the COD set alone. Without those rows, the reported gains could come mostly from dropping 61.7% of DUTS (data curation) plus the extra DLM parameters, not from joint supervision. The stress-test note is right on this. The reader's oracle concern is secondary but real: the ranking score in Eq. (7) is the sum of the exact metrics they report, and the oracle is the authors' own SENet. That said, the Top-K vs Random-K vs Bottom-K comparison shows selection direction matters, so the effect is not pure circularity. Also, all numbers are single-seed; the SOD/COD improvements are small in places, and we need variance to know if they are noise.\n\nNone of this kills the paper. The controlled comparisons on a common backbone (Table V, and the data-ratio sweep) are trustworthy, and the recipe itself is valuable even if the \"joint learning\" framing is overstated. This is a serious empirical paper for anyone working on multi-task segmentation or training-set curation. Deserves peer review, but the final version should add the independent-training controls and multi-seed results before the causal wording can be accepted.","headline":"A well-ablated system paper whose headline claim—that joint learning benefits both SOD and COD—is confounded by a missing control: the final model changes architecture, data subset, and training regime simultaneously.","tokens_in":23116,"tokens_out":3213,"would_cite":false,"duration_ms":38523,"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":"A single fully shared network can learn both salient and camouflaged object detection, and each task improves the other.","keywords":["salient object detection","camouflaged object detection","joint learning","multi-task learning","distribution learning module","saliency-based sampling","vision transformer","binary segmentation"],"falsifier":"Re-run the Table I ablations with SBSS ranking produced by a different pretrained SOD model or by human-annotated pair quality instead of SENet's agreement, and compare top-K versus bottom-K selection. If the gap between Top-K and Bottom-K in Table III disappears or reverses when the scorer changes, the SBSS gains are a selection artifact of that particular scorer and its overlap with the test benchmarks; if the gap persists, the pruning is selecting genuinely better training supervision.","tokens_in":21874,"feed_emoji":"🎯","tokens_out":9122,"duration_ms":91780,"temperature":0.7,"pith_summary":"This paper claims that salient object detection and camouflaged object detection — two segmentation tasks long thought to conflict in a shared network — can be learned together in a way that makes each task improve the other. The mechanism is SCJoint: one fully shared transformer network, with the only task-specific parts being two tiny distribution learning modules in the decoder, each a learned mean and variance applied to decoder features. To handle the fact that the SOD training set is about 2.5 times larger than the COD set, the authors add SBSS, which ranks 10,553 SOD image-mask pairs by how well a pretrained model's prediction matches the ground truth and keeps only the top 4,040 pairs. The trained network, JoNet, beats its own independently trained baseline on both tasks and reports state-of-the-art results on nine SOD and COD benchmarks. If true, this overturns the standard assumption that opposite tasks must be kept in separate specialist models.","feed_headline":"One shared network beats separate specialists on both tasks","feed_subtitle":"Adding about 2,000 task-specific parameters lets one model beat separate experts on both tasks.","key_machinery":"The load-bearing mechanism is the Distribution Learning Module (DLM), a per-task affine normalization inserted into every Transformer block of the decoder: for task $t$, $X' = (X - \\mu_t)/\\sqrt{\\sigma_t^2}$, with $(\\mu_t, \\sigma_t)$ learned only from that task's mini-batches. This is the entire task-specific parameter budget — about 2K parameters — and it is what lets one shared network keep two decoding modes. The second mechanism is the Saliency-Based Sampling Strategy (SBSS), which ranks image-mask pairs by the agreement between a pretrained SOD model's prediction and the ground-truth mask, using the composite $S = S_\\alpha + E_\\phi + F_\\beta + (1-M)$, and keeps the top 4,040 DUTS pairs.","core_discovery":"The paper's central claim is that SOD and COD share a unified encoding distribution but have different decoding distributions, so a single network can serve both tasks if each task's decoder features are transformed by its own learnable mean and variance. SCJoint inserts a DLM into every Transformer block of the decoder; the shared parameters are updated by both tasks while $(\\mu_s, \\sigma_s)$ and $(\\mu_c, \\sigma_c)$ are updated only by their own task's loss, decoupling \"salient\" from \"camouflaged\" for roughly 2K additional parameters. SBSS addresses the data imbalance by pruning DUTS from 10,553 to 4,040 image-mask pairs, using the composite score $S = S_\\alpha + E_\\phi + F_\\beta + (1-M)$ t","pith_inferences":["The SBSS ranking score is the same composite of the four evaluation metrics used for benchmarking, so the pruning criterion is partly aligned with the benchmark itself; a sharper test of \"quality\" would rank pairs with an independent model or human annotation and see whether the gains persist.","If the shared-encoder claim is general, DLM should transfer to other pairs of binary segmentation tasks that share low-level features but differ in what they attend to; the shadow-detection result already suggests this, and testing on, say, edge detection plus salient objects would be a natural next step.","Because DLM is mathematically a per-task scale-and-shift on shared decoder features, a plausible simplification — worth testing — is that the same benefit comes from per-task affine transforms applied only to the final decoder features, rather than in every Transformer block.","JoNet is initialized from MAE, not from a jointly pretrained model, so the reported gains are attributable to the training scheme itself; this makes the method a candidate for use with other transformer backbones, and one could test whether the gains scale with backbone strength."],"forward_implications":["A fully shared network with roughly 2K task-specific parameters can match or beat separate specialist models on both SOD and COD, so a single deployed model can replace two task-specific ones.","Joint training with SCJoint plus SBSS improves both tasks over independent training — DUTS $S_\\alpha$ rises .921 to .931 and CAMO $S_\\alpha$ .875 to .889 in the ablations — not just one task at the other's expense.","The method locates the task conflict in the decoder: inserting DLM in the encoder instead of the decoder sharply degrades SOD (Table II), supporting the paper's claim that the two tasks share encoding and differ in decoding.","Pruning 61.7% of the SOD training set improves SOD results, implying that the discarded low-agreement pairs were providing misleading supervision for joint learning.","SCJoint extends to at least one third binary segmentation task: adding shadow detection to SOD and COD improves all three tasks over independent training (Table IX), indicating the scheme is not specific to the salient/camouflaged pair."],"supporting_citations":[{"why":"Supplies the SENet backbone on which JoNet is built, the pretrained model used to score DUTS pairs in SBSS, and the independent-training baseline (IT) that all joint-learning gains are measured against.","marker":"[3]"},{"why":"An earlier SOD+COD joint-learning method the paper compares against; it represents the shared-decoder strategy and is cited as evidence that naive joint learning underperforms.","marker":"[1]"},{"why":"Follow-up uncertainty-aware joint SOD+COD work cited for the claim that COD samples behave as harder SOD samples and for the risk that the larger SOD set dominates training.","marker":"[2]"},{"why":"AdvProp supplies the conceptual template: separate batch-norm-style statistics for two data distributions, which SCJoint adapts to SOD and COD decoder features.","marker":"[16]"},{"why":"Provides the pixel-position-aware weighted BCE+IoU loss used for both tasks (Eq. 6), keeping the objective identical across tasks.","marker":"[28]"},{"why":"MAE pretraining initializes both encoder and decoder of JoNet and the SENet baseline, so the comparison isolates the joint-training scheme rather than the initialization.","marker":"[62]"},{"why":"DUTS is the SOD training set that SBSS prunes from 10,553 to 4,040 pairs, and it is the primary SOD benchmark in the ablations.","marker":"[64]"},{"why":"CAMO provides part of the COD training set (with COD10K) and is the COD benchmark where the paper reports the largest joint-learning gains.","marker":"[66]"},{"why":"Introduces the COD task and the COD10K dataset, supplying both a training split and a test benchmark, and its SINet model serves as a comparison method.","marker":"[5]"}],"fun_headline_variants":["2K params let one net master both salient and camouflaged detection","Shared decoder, 2K task params: one model beats two specialists","Contrary to belief, joint training improves both SOD and COD","Minimal task-specific params unify salient and camouflaged detection","One network, 2K extra parameters: both detection tasks, one winner"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing assumption is that a pretrained saliency model's agreement with a ground-truth mask measures the quality of that image-mask pair, so pruning away the 61.7% lowest-agreement DUTS pairs improves the training data rather than merely selecting easy or test-like samples.","fun_headline_variants_meta":{"raw":{"variants":["2K params let one net master both salient and camouflaged detection","Shared decoder, 2K task params: one model beats two specialists","Contrary to belief, joint training improves both SOD and COD","Minimal task-specific params unify salient and camouflaged detection","One network, 2K extra parameters: both detection tasks, one winner"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001406,"raw_usage":{"total_tokens":5587,"prompt_tokens":879,"completion_tokens":4708,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":4614}},"tokens_in":623,"tokens_out":4708,"duration_ms":31784,"temperature":1.0,"reasoning_tokens":4614,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:59:02.201098+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Table I ablations with SBSS ranking produced by a different pretrained SOD model or by human-annotated pair quality instead of SENet's agreement, and compare top-K versus bottom-K selection. If the gap between Top-K and Bottom-K in Table III disappears or reverses when the scorer changes, the SBSS gains are a selection artifact of that particular scorer and its overlap with the test benchmarks; if the gap persists, the pruning is selecting genuinely better training supervision.","supporting_citations":[{"cited_title":"A Simple yet Effective Network based on Vision Transformer for Camouflaged Object and Salient Object Detection","cited_arxiv_id":"2402.18922","evidence_quote":"Supplies the SENet backbone on which JoNet is built, the pretrained model used to score DUTS pairs in SBSS, and the independent-training baseline (IT) that all joint-learning gains are measured against."},{"cited_title":"Joint Salient Object Detection and Camouflaged Object Detection via Uncertainty-aware Learning","cited_arxiv_id":"2307.04651","evidence_quote":"Follow-up uncertainty-aware joint SOD+COD work cited for the claim that COD samples behave as harder SOD samples and for the risk that the larger SOD set dominates training."},{"cited_title":"Adversarial examples improve image recognition,","cited_arxiv_id":null,"evidence_quote":"AdvProp supplies the conceptual template: separate batch-norm-style statistics for two data distributions, which SCJoint adapts to SOD and COD decoder features."},{"cited_title":"F 3net: fusion, feedback and focus for salient object detection,","cited_arxiv_id":null,"evidence_quote":"Provides the pixel-position-aware weighted BCE+IoU loss used for both tasks (Eq. 6), keeping the objective identical across tasks."},{"cited_title":"Masked au- toencoders are scalable vision learners,","cited_arxiv_id":null,"evidence_quote":"MAE pretraining initializes both encoder and decoder of JoNet and the SENet baseline, so the comparison isolates the joint-training scheme rather than the initialization."},{"cited_title":"Anabranch network for camouflaged object segmentation,","cited_arxiv_id":null,"evidence_quote":"CAMO provides part of the COD training set (with COD10K) and is the COD benchmark where the paper reports the largest joint-learning gains."}],"review_version":1}