{"id":"28d2bcbb-0845-4430-bf7f-1b3809562cee","arxiv_id":"2411.14201","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"RASM, a lightweight shadow-removal network with regional attention that aggregates neighboring non-shadow information, reports SOTA PSNR on ISTD+ and competitive accuracy on SRD with low FLOPs.","lead":"This paper introduces RASM, a lightweight shadow-removal network that uses a regional attention mechanism to let each shadowed area borrow information from its immediate non-shadowed surroundings. The method reports top accuracy on the ISTD+ benchmark while using fewer parameters and FLOPs than most prior attention-based models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ISTD+ results are tuned on the same test set used for comparison; the 0.90 dB SOTA margin may reflect selection bias rather than genuine improvement.","rationale":"The reader's verdict is CONDITIONAL with medium correctness risk, which matches my assessment. However, the reader's weakest_assumption focuses on the regional-attention hypothesis (that proximate non-shadow context is the key information source), whereas I find the more load-bearing concern to be the test-set-based hyperparameter selection and missing error bars, which directly undermine the quantitative SOTA claim. The hypothesis is motivational; even if it were imperfect, the method's empirical performance is the actual contribution. The test-set tuning is visible in Table 5, where r and d are selected by comparing PSNR/RMSE on ISTD+, the same set used for the final comparison in Table 1. This is a classic selection-on-test problem. The lack of error bars compounds it: the reported 0.90 dB improvement could be within run-to-run variance. I therefore partially agree with the reader's overall conditional verdict but identify a different core concern. My concrete test settles whether the ISTD+ margin is genuine; if the margin survives held-out validation, the central claim stands, and if it does not, the paper's main quantitative contribution is unsupported. The SRD overclaim is secondary and explicitly acknowledged in the paper's own text, so it is not the primary concern.","tokens_in":11980,"tokens_out":5028,"duration_ms":46935,"concrete_test":"Split the ISTD+ test set (540 images) into a validation split and a held-out test split, e.g., 50/50 or 5-fold cross-validation. Retune region size r and dilation d on the validation split only, then retrain RASM with the selected hyperparameters and evaluate on the held-out split. If the PSNR advantage over ShadowFormer/ShadowDiffusion is less than 0.90 dB or reverses, the SOTA claim is not robust. Additionally, run three independent training seeds with fixed hyperparameters and report mean +/- std on both ISTD+ and SRD to check whether the margin exceeds run-to-run noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim, that RASM surpasses SOTA by 0.90 dB PSNR on ISTD+, rests on results in Table 1, but Table 5 shows the regional-attention hyperparameters (region size r=11, dilation d=2) were chosen by evaluating variants on the same ISTD+ test set. No error bars or multi-seed results are reported for any configuration, so the reported 40.73 dB is the maximum of several test-set evaluations, an optimistically biased estimate of expected performance. Without a held-out validation split or significance testing, the margin over ShadowFormer (39.67) and ShadowDiffusion (39.82) may not survive proper evaluation. This directly threatens the 'superior performance' claim, because the efficiency comparison alone does not establish superiority: BMNet has fewer parameters (0.4M) and GFLOPs (11.6G) than RASM (5.2M, 25.2G).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a lightweight shadow removal model, RASM, built around a regional attention mechanism that lets each shadowed token attend to its surrounding non-shadowed tokens, with a dilation option to enlarge the receptive field. The architecture is a U-shaped encoder-decoder with channel attention and a Regional Attention Module at the bottleneck, trained with Charbonnier and VGG-based perceptual losses. The authors evaluate on ISTD+ and SRD, reporting state-of-the-art PSNR/SSIM/RMSE on ISTD+ and strong results on SRD, with 5.2M parameters and 25.2 GFLOPs. The central claim, stated in the abstract, is that the method delivers superior performance over other state-of-the-art models in both accuracy and efficiency.","tokens_in":12209,"tokens_out":3550,"duration_ms":34327,"significance":"If the empirical claims hold, the paper makes a useful contribution: it identifies a simple inductive bias for shadow removal—information from adjacent non-shadow regions is especially informative—and packages it in a lightweight, efficient network. The strengths include publicly available code, comparisons on two widely used datasets with a consistent evaluation protocol, and a clear ablation isolating the regional attention mechanism versus window attention. The regional attention design is plausible and the qualitative attention visualizations support the mechanism's behavior. However, the headline \"superior performance\" claim is currently not fully supported: the ISTD+ results appear to be selected through test-set hyperparameter tuning, no uncertainty measures are reported, and on SRD the shadow-region accuracy trails two prior methods. The significance of the contribution is therefore conditional on additional validation that the reported margins are not artifacts of selection bias.","major_comments":[{"comment":"The final model configuration (region size r=11, dilation d=2) is chosen by evaluating variants on the ISTD+ test set, and the same test set is used to report the headline results in Tab. 1. The reported 40.73 dB PSNR is thus a selected maximum over multiple test-set evaluations, which introduces optimistic bias and can inflate the margin over ShadowFormer and ShadowDiffusion. Please either select hyperparameters on a held-out validation split (or via cross-validation) and then evaluate a single configuration on the test set, or report the selection procedure explicitly and provide significance testing, such as paired bootstrap confidence intervals over images, for the key comparisons in Tab. 1.","section":"Sec. 4.3, Tab. 5"},{"comment":"The abstract and introduction claim superior performance over state-of-the-art models, but on the SRD dataset the shadow-region PSNR of RASM (37.91 dB) is below ShadowDiffusion (38.72 dB) and Li et al. (39.33 dB), and the paper's own text acknowledges this gap. The superiority claim is therefore dataset- and metric-dependent; please scope the claim to the settings in which it holds, or provide a well-defined multi-objective criterion (e.g., Pareto dominance over accuracy and efficiency) under which RASM is superior on both datasets.","section":"Sec. 4.2.1, Tab. 2"},{"comment":"The efficiency comparison does not fully support the claim of \"superior performance in terms of accuracy and efficiency\": BMNet has 0.4M parameters and 11.6 GFLOPs, whereas RASM has 5.2M parameters and 25.2 GFLOPs, so RASM is not the most efficient among the compared methods. The efficiency advantage is relative mainly to large diffusion and transformer models. Please present a Pareto-style analysis of accuracy versus FLOPs/parameters, or explicitly characterize the accuracy-efficiency trade-off that RASM achieves, rather than stating a general efficiency superiority.","section":"Sec. 4.2.1, Tab. 3"},{"comment":"No error bars, standard deviations, or multiple-seed results are reported for any configuration. Some of the reported margins are very small in absolute terms (e.g., non-shadow PSNR 39.23 vs. 39.15 for Liu et al. on ISTD+; SSIM 0.993 vs. 0.992), so without uncertainty quantification the reader cannot assess whether these differences are statistically meaningful. Please report at least three independent training runs with mean and standard deviation for the main comparisons, or use paired statistical tests if single-seed evaluation is standard for the dataset.","section":"Sec. 4.2, Tabs. 1 and 2"}],"minor_comments":[{"comment":"Equation (5) is ambiguous about the scope of the softmax: it should read softmax(A_i^r / sqrt(d)) V_i^r, with parentheses around the softmax argument, rather than softmax(A_i^r / sqrt(d) V_i^r).","section":"Sec. 3.2.2, Eq. (5)"},{"comment":"The text states \"AdamW optimizer with the momentum as (0.9, 0.999)\"; AdamW uses betas, not momentum, so please rename this to betas for accuracy.","section":"Sec. 4.1"},{"comment":"In the concluding remarks, \"ISRD+\" should be \"ISTD+\".","section":"Sec. 5"},{"comment":"The sentence \"The competitors are consist of 11 methods\" is ungrammatical; please change to \"The competitors consist of 11 methods.\"","section":"Sec. 4.2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a well-executed systems paper with a clear and plausible idea, and the lack of error bars is common in this subfield. The main concern is that the headline ISTD+ result is selected from test-set tuning, which is a methodological issue the authors can likely address with a validation split and significance testing rather than a fundamental flaw. I would support publication after the central claims are re-scoped or the validation is strengthened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a straightforward, honest application of neighborhood and dilated attention to shadow removal. The authors adapt NAT/DiNAT, say so clearly, and wrap it in a small U-Net with channel attention. The central hypothesis — that adjacent non-shadow context matters more than global context for recovering a shadowed region — is plausible, and the window-versus-regional ablation supports it. The experiments are extensive, comparisons use published numbers, and the SRD limitation is admitted outright (shadow-region PSNR trails ShadowDiffusion and Li et al.). The code appears to be released, which counts for something.\n\nThe soft spots are real but not fatal. The biggest one is Table 5: the region size and dilation are chosen by evaluating on the ISTD+ test set, and the same test set produces the headline 40.73 dB. That is textbook selection bias, and the 0.90 dB margin over ShadowFormer and ShadowDiffusion is very likely optimistic. No error bars or multi-seed runs are reported anywhere, so we cannot tell whether 40.73 is meaningfully above 39.82. The stress-test note got this right, and it is the main thing I would push back on.\n\nThe abstract's \"superior performance in terms of accuracy and efficiency\" is also a bit loose. RASM is more accurate than BMNet, but BMNet has 0.4M parameters and 11.6 GFLOPs against RASM's 5.2M and 25.2 GFLOPs. RASM is not the most efficient; it is the most accurate at moderate cost. That is a fine trade-off, but the claim should say so.\n\nWho gets value from this? People working on shadow removal will find a useful lightweight baseline and a clear description of the design. People interested in local-attention mechanisms for partial-corruption restoration will also benefit, even though the novelty over NAT/DiNAT is incremental. The methodology lesson about test-set tuning is a good reading-group discussion point.\n\nMy recommendation: yes, this deserves peer review. It is a competent, publishable paper; the flaws are addressable. A serious referee should ask for a validation split or repeated runs, and a more careful efficiency claim. The central design choice is credible, and there is enough reproducibility (code, detailed ablations) to warrant referee time rather than a desk reject.","headline":"A workmanlike shadow-removal paper with a genuine efficiency-accuracy trade-off, undermined a bit by test-set tuning and missing error bars.","tokens_in":12698,"tokens_out":2067,"would_cite":false,"duration_ms":21350,"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":"Shadow removal works best when each shadow region reads its immediate non-shadow neighbors, and a 5.2M-parameter network implementing this idea reaches 40.73 dB PSNR on ISTD+ shadow regions, beating all prior models with far less compute.","keywords":["shadow removal","regional attention","lightweight network","transformer","ISTD+","SRD","image restoration","neighborhood attention"],"falsifier":"Generate paired shadow/shadow-free images in which the same local neighborhood around a shadow is rendered under two different distant lighting conditions, so the correct unshadowed appearance differs while the nearby context is identical. If RASM cannot reproduce the two different ground truths, the proximity assumption is false. A complementary test: in ISTD+ test images, corrupt the immediate non-shadow surround and measure how much RASM's shadow-region PSNR drops; a small drop would indicate that nearby context is not actually the deciding signal.","tokens_in":11803,"feed_emoji":"☀️","tokens_out":8204,"duration_ms":67992,"temperature":0.7,"pith_summary":"The paper is trying to establish that the decisive information for removing a shadow lives in the non-shadowed pixels immediately around it, and that a model built on this idea can be both more accurate and much lighter than existing attention- or diffusion-based shadow removers. To exploit that premise, it proposes regional attention: each shadow-location token is allowed to read from a fixed-size surrounding region rather than from the whole image or a rigid window. The reported results back the claim: on ISTD+ the network reaches 40.73 dB PSNR in shadow regions, 0.90 dB above the prior best, while using 5.2M parameters and 25.2 GFLOPs. If the premise holds, accurate shadow removal no longer requires heavy global models, which matters for mobile and downstream vision applications.","feed_headline":"Neighbor-focused attention lifts shadow removal to 40.73 dB on ISTD+","feed_subtitle":"A 5.2M-parameter model beats prior art by 0.90 dB on shadow regions at a fraction of the compute.","key_machinery":"The regional attention module (RAM) carries the argument: for each token $i$ on the feature map, it computes dot-product attention between the query $Q_i$ and the keys of the $r$ tokens in a surrounding region, adds a relative position bias, and applies a softmax-scaled weighted sum over the corresponding values (Eqs. 3-5). With region size 11 and dilation factor 2, the effective receptive field expands without increasing parameters or FLOPs, following the dilated-neighborhood idea. RAM replaces the window-attention or self-attention blocks used by prior shadow-removal transformers, and is paired with a Squeeze-and-Excitation-style channel attention module so that global channel statistics are still captured cheaply. The loss is a weighted combination of Charbonnier content loss and VGG-19 perceptual loss.","core_discovery":"The central discovery is that shadow restoration is best framed as a region-context problem, not a global-context problem. RASM treats the shadow mask as a guide and lets every shadowed token attend to the tokens in its immediate neighborhood using a regional attention module with region size 11 and dilation 2; the surrounding non-shadow information is what reconstructs the shadowed pixels. On ISTD+, this yields 40.73 dB PSNR and 0.993 SSIM in the shadow region, surpassing the previous SOTA by 0.90 dB and also leading in all-image PSNR (36.16 dB) and RMSE (2.53). On SRD, the model achieves the best all-image RMSE (3.37) and the best non-shadow-region metrics, though its shadow-region PSNR (37.91 dB) trails two competitors, which the authors attribute to stricter adherence to the predicted shadow mask. The same network is an order of magnitude smaller than diffusion-based alternatives, using 5.2M parameters and 25.2 GFLOPs, so the paper's claim is that accuracy and efficiency are jointly achievable once attention is aligned with the geometry of shadows.","pith_inferences":["The proximity hypothesis could be tested directly: corrupt or erase the immediate non-shadow surround in ISTD+ test images and measure the PSNR drop; a large drop would confirm that nearby context is the load-bearing signal, while a small drop would suggest global cues do more work than the paper assumes.","Because the method needs a shadow mask at inference, its practical reach depends on mask quality; a mask-free variant or joint detection module would be a natural extension the paper does not address.","The efficiency claim is metered at 256x256; at higher resolutions the fixed 11x11 region covers a smaller fraction of the scene, so resolution scaling behaviour is an open question.","Regional attention may transfer to other partial-corruption tasks such as reflection removal or object inpainting, since those share the property that the correct output is locally inferable from clean surroundings."],"forward_implications":["If regional attention is the right inductive bias, shadow-removal systems can drop global self-attention and diffusion components entirely, making real-time on-device shadow removal practical.","The ISTD+ numbers imply that a 5.2M-parameter model can outperform a 55.2M-parameter diffusion model by 0.90 dB in shadow regions, so the task does not require large models once the attention geometry matches the corruption pattern.","The window-vs-regional ablation (40.06 vs 40.73 dB) shows the gain comes from aligning attention with shadow boundaries, not from added capacity, predicting that other mask-guided restoration tasks will benefit from the same substitution.","The dilation experiment shows receptive-field range can be enlarged at zero FLOPs cost by increasing dilation, decoupling how far the model looks from how much it computes."],"supporting_citations":[{"why":"ShadowFormer: the window-attention shadow-removal baseline whose results and attention design RASM directly compares against.","marker":"[10]"},{"why":"Neighborhood Attention Transformer: supplies the local-attention operation that the regional attention module adapts to shadow removal.","marker":"[14]"},{"why":"Dilated Neighborhood Attention Transformer: provides the dilation technique for enlarging receptive field without added parameters.","marker":"[13]"},{"why":"The adjusted ISTD+ dataset: the main benchmark where RASM reports its SOTA shadow-region PSNR of 40.73 dB.","marker":"[23]"},{"why":"SRD dataset: the large paired shadow-removal benchmark used for the efficiency and all-image RMSE comparison.","marker":"[31]"},{"why":"DHAN: provides the predicted shadow masks used for SRD evaluation, and is itself a strong baseline in the comparisons.","marker":"[2]"},{"why":"ShadowDiffusion: the prior state of the art whose ISTD+ shadow-region score (39.82 dB) RASM claims to surpass by 0.90 dB.","marker":"[11]"},{"why":"BMNet: the most parameter-efficient prior model (0.4M parameters) that anchors the efficiency comparison.","marker":"[38]"}],"fun_headline_variants":["Regional attention makes shadow removal fast and accurate","40.73 dB on ISTD+ with a 5.2M-parameter shadow remover","Neighborhood-focused attention beats global for shadows","Shadow removal: local context wins, model is tiny","RASM: 0.90 dB better, 10x smaller than diffusion rivals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model rests on the premise that the information needed to restore a shadowed area is concentrated in the non-shadowed pixels immediately surrounding it and becomes less important with distance; if shadow appearance is governed by global illumination or distant scene context, regional attention will miss the deciding cues.","fun_headline_variants_meta":{"raw":{"variants":["Regional attention makes shadow removal fast and accurate","40.73 dB on ISTD+ with a 5.2M-parameter shadow remover","Neighborhood-focused attention beats global for shadows","Shadow removal: local context wins, model is tiny","RASM: 0.90 dB better, 10x smaller than diffusion rivals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000292,"raw_usage":{"total_tokens":1726,"prompt_tokens":992,"completion_tokens":734,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":644}},"tokens_in":608,"tokens_out":734,"duration_ms":6446,"temperature":1.0,"reasoning_tokens":644,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:24:53.858356+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate paired shadow/shadow-free images in which the same local neighborhood around a shadow is rendered under two different distant lighting conditions, so the correct unshadowed appearance differs while the nearby context is identical. If RASM cannot reproduce the two different ground truths, the proximity assumption is false. A complementary test: in ISTD+ test images, corrupt the immediate non-shadow surround and measure how much RASM's shadow-region PSNR drops; a small drop would indicate that nearby context is not actually the deciding signal.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ShadowFormer: the window-attention shadow-removal baseline whose results and attention design RASM directly compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Neighborhood Attention Transformer: supplies the local-attention operation that the regional attention module adapts to shadow removal."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SRD dataset: the large paired shadow-removal benchmark used for the efficiency and all-image RMSE comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DHAN: provides the predicted shadow masks used for SRD evaluation, and is itself a strong baseline in the comparisons."}],"review_version":1}