{"id":"6bef2337-e4ca-46f8-b6bd-3a7ca6c58378","arxiv_id":"2501.07984","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"TANet replaces pixel-pair self-attention with attention over threshold-quantized pixel value groups, achieving state-of-the-art segmentation on ISPRS Vaihingen and Potsdam.","lead":"Remote sensing segmentation is expensive when every pixel is compared with every other pixel. This paper groups pixels into value intervals and computes attention between those groups, reporting faster, more accurate labeling on two aerial benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA accuracy claim is confounded by TTA/OHEM/aux loss; without them TANet trails several cited baselines","rationale":"The reader's verdict is CONDITIONAL and already flags the TTA/OHEM fairness concern in its rationale, which aligns with my analysis. However, the reader's explicitly identified weakest assumption is the semantic grouping premise at Eq. (1) and Eq. (7). I disagree that this is the most load-bearing point. The grouping premise is an inductive bias whose validity can be evaluated empirically; the internal ablation (Table VII) and the visualizations provide some evidence that the mechanism helps in the tested configuration. The fairness of the SOTA comparison, by contrast, directly determines whether the headline claim 'TANet outperforms other methods' is true at all. The paper's own Table X quantifies that OHEM, auxiliary loss, and TTA account for a 1.1 mIoU gain on Vaihingen, which is larger than the gap between TANet and several top baselines. Since the paper does not disclose the training/evaluation protocol of the baselines, the SOTA comparison is not controlled, and the central claim could be an artifact of test-time augmentation and hard-example mining rather than the proposed threshold attention mechanism. This does not invalidate the method entirely; it means the empirical support is currently insufficient. A conditional acceptance requiring the authors to re-run a controlled comparison or to report the baselines' settings is the appropriate disposition, so I keep the reader's verdict unchanged. The efficiency contradiction (49 GFLOPs vs slower inference than SA) is a secondary concern worth mentioning but does not supersede the accuracy-comparison confound, because the paper's own efficiency table is framed at module level while the inference table is end-to-end and includes backbone overhead.","tokens_in":16466,"tokens_out":5046,"duration_ms":50220,"concrete_test":"Re-run TANet on the Vaihingen test set with the configuration of the first row of Table X (no OHEM, no auxiliary loss, no TTA) and compare its mIoU, OA, and Mean F1 directly against the published scores of HMANet, MFNet, CTMFNet, and HBCNet in Table XII under the same data split and evaluation protocol. If TANet's no-trick scores do not exceed these baselines, the conclusion that TANet 'outperforms other methods' is not supported by TAM alone and must be restated or the baselines must be re-run under identical conditions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that TANet outperforms state-of-the-art on Vaihingen and Potsdam. This claim rests on Tables XII and XIII, yet the reported TANet uses OHEM, auxiliary loss, and test-time augmentation. Table X shows these additions add +1.10 mIoU on Vaihingen and +0.68 on Potsdam. Without them, TANet's mIoU on Vaihingen is 83.35, which is below HMANet (83.49), MFNet (83.50), CTMFNet (84.34), and HBCNet (84.21) listed in the same comparison table (Table XII). The paper does not state which baseline methods used TTA, OHEM, or auxiliary losses. If those baselines did not use such tricks, the comparison is not controlled and the accuracy advantage attributed to TAM is not established. In addition, the efficiency claim is internally ambiguous: Table VIII reports 49 GFLOPs for the proposed modules versus 619 for SA, but Table IX shows TANet's end-to-end inference time (0.091s) is slower than SA (0.076s). The uncontrolled accuracy comparison is the more load-bearing issue because the paper's strongest claim depends on the unadjusted Table XII/XIII results, while the internal ablation in Table VII only shows a modest TAM benefit over the backbone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Threshold Attention Mechanism (TAM) that quantizes each channel of a feature map into L intervals, computes attention between the resulting L per-channel threshold features, and then reassigns the attention weights to pixels via a one-hot position matrix, achieving complexity that is linear in the number of pixels. Based on TAM, the authors build TANet, consisting of an Attentional Feature Enhancement Module (AFEM) for shallow features and a Threshold Attention Pyramid Pooling module (TAPP) for deep features. The method is evaluated on the ISPRS Vaihingen and Potsdam semantic segmentation benchmarks, reporting state-of-the-art or competitive mIoU, OA, and F1 scores.","tokens_in":16811,"tokens_out":8684,"duration_ms":79341,"significance":"If the claims are substantiated, the threshold attention idea is a potentially useful alternative to dense self-attention for high-resolution remote sensing segmentation, and the paper includes a reasonably extensive experimental study with parameter ablations and module-level efficiency comparisons. The manuscript also makes its core formulas explicit, which is helpful for reproducibility. However, the two headline claims — that TAM 'significantly reduces computational effort' and that TANet outperforms state-of-the-art methods — are currently undermined by internal inconsistencies and uncontrolled comparisons, as detailed below.","major_comments":[{"comment":"The state-of-the-art comparison is not controlled for training enhancements. The TANet entries in Tables XII and XIII use OHEM, auxiliary loss, and test-time augmentation; Table X shows these add +1.10 mIoU on Vaihingen (83.35→84.45). Without them, TANet's Vaihingen mIoU of 83.35 is below HMANet (83.49), MFNet (83.50), CTMFNet (84.34), and HBCNet (84.21) in the same table. The paper never states whether the cited baselines were trained or evaluated with these techniques. Please provide a controlled comparison, such as reporting TANet without OHEM/aux loss/TTA as the primary SOTA number, or reporting baselines with and without the same enhancements.","section":"§V-C, Tables X and XII"},{"comment":"The computational-complexity claim is internally inconsistent. Table VIII reports AFEM+TAPP at 49 GFLOPs versus 619 for standard self-attention, but Table IX shows TANet's end-to-end inference time (0.091 s) is slower than SA (0.076 s). The paper should clarify that the GFLOPs comparison is at the module level only, report end-to-end FLOPs and inference time for the full networks, and discuss why the module-level reduction does not translate into faster inference.","section":"§V-B2, Tables VIII-IX and Abstract"},{"comment":"The discretization formula uses a factor 2L, whereas Eq. (1) defines L threshold levels. As written, P can take 2L+1 distinct values (0,...,2L), but Eq. (8) requires P'_c ∈ R^{L×N} (L rows) for the one-hot encoding. This inconsistency breaks the TAM computation as described. Please correct Eq. (7) to use L (or adjust the one-hot dimension) and state the exact value range of P.","section":"§III-B, Eq. (7)"},{"comment":"The Vaihingen data split description is self-contradictory. The text says 16 training and 17 test regions, then lists 17 image IDs for training, and then says 'Thirty labeled images are used for validation,' which cannot be reconciled with a 33-image dataset. Please provide an exact, consistent split of training, validation, and test images.","section":"§IV-A1"},{"comment":"The OHEM parameter S is set to 10,000 but is never defined, and the description of OHEM via a probability threshold θ is nonstandard (OHEM typically selects a fixed number of hardest examples). Please define S and describe the hard-example selection procedure precisely.","section":"§III-E"},{"comment":"The core premise that per-channel intensity intervals correspond to semantically coherent pixel regions is not analyzed or compared with alternative grouping schemes such as adaptive binning, k-means, or superpixels. Because this premise underlies the claimed accuracy-efficiency advantage of TAM, please provide at least an ablation or analysis that isolates the effect of the quantization/grouping strategy.","section":"§III-B"}],"minor_comments":[{"comment":"The caption spells 'Vaihingen' as 'VAHINGEN'; please correct it.","section":"Table X caption"},{"comment":"In the sentence 'the 15 images labeled as follows:', the list contains 17 IDs; please correct the count and the surrounding text.","section":"§IV-A1"},{"comment":"The term 'Test-Time Enhancement' should be 'Test-Time Augmentation' throughout the paper.","section":"§V-C"},{"comment":"The AdamW optimizer's weight decay value is not reported; please state it for reproducibility.","section":"§IV-C"},{"comment":"Please state explicitly whether the L1 and L2 parameters were selected on the validation set or on the test set, so readers can assess the parameter-study protocol.","section":"§V-A"},{"comment":"In the CASIA3 row, the Building F1 score of 86.80 is anomalously low compared with the other rows; please verify it against the original source.","section":"Table XIII"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the uncontrolled state-of-the-art comparison: the reported TANet numbers include OHEM, auxiliary loss, and test-time augmentation, and the ablation in Table X shows these are worth more than 1 mIoU. The efficiency claim is also hard to accept as stated given that TANet is slower end-to-end than plain self-attention (Table IX). The Eq. (7) inconsistency is a concrete technical error that suggests the manuscript needs careful proofreading. If the authors can provide controlled comparisons, fix the formula and split descriptions, and clarify the efficiency story, the contribution could be publishable after a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here is my read on arXiv:2501.07984. The genuinely new thing is threshold attention: quantize each channel into L intervals, compute softmax attention between the L interval representatives, then reassign the attended weights to pixels by one-hot lookup. That is a real construction, distinct from the external attention and kernel attention they cite, and it gives linear complexity in pixel count. The modules (AFEM, TAPP) are reasonable, and the internal ablation in Table VII shows the TAM module alone beats the ResNet baseline by 1.21 mIoU on Vaihingen. That is the strongest evidence in the paper.\n\nThe soft spots are real and load-bearing. The SOTA comparison in Tables XII/XIII is not controlled. The final model includes OHEM, auxiliary loss, and test-time augmentation. Table X shows those add 1.10 mIoU on Vaihingen (83.35 to 84.45). Without them, TANet's 83.35 is below HMANet, MFNet, CTMFNet, and HBCNet in the same table. The paper never says which baselines used TTA/OHEM/aux loss, so the claimed accuracy advantage over the state of the art is not established. That is a problem, because the abstract's main claim rests on those tables. Second, the efficiency story is internally odd: Table VIII reports 49 GFLOPs for the two modules versus 619 for self-attention, but Table IX shows TANet end-to-end inference is slower than the SA model (0.091 vs 0.076 s). The paper admits the time cost is not optimal, but the abstract's 'significantly reduces computational effort' needs qualification. Third, there are clarity issues that should have been caught: the Vaihingen split says 15 images but lists 17, then says 30 images are used for validation, which cannot be right for a 33-image dataset. The loss formulation for OHEM is not specified, and a parameter S=10,000 is introduced without definition. The threshold levels L1/L2 are tuned on the evaluation datasets, or at least not clearly on a separate validation split. No code is released, which makes the comparison hard to audit. Finally, the deeper conceptual premise — that pixels in the same per-channel value interval share enough semantics that attention between intervals is a faithful replacement for pixel-pair attention — is not analyzed and is only supported by the benchmark numbers.\n\nNone of these are fatal to the idea. The mechanism is genuinely new and the gains over the backbone are credible. But the central SOTA claim is currently confounded, and the paper needs a controlled comparison, code or explicit baseline settings, and a fix to the efficiency and dataset descriptions. I would send it to review, not desk reject; the idea deserves a serious look, and the authors have done enough work that a revision could resolve the concerns.","headline":"Threshold attention is a genuinely new mechanism, but the SOTA comparison is confounded by TTA/OHEM/aux loss and the efficiency claim needs qualification; worth review, not desk reject.","tokens_in":17295,"tokens_out":2788,"would_cite":false,"duration_ms":26075,"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":"This paper claims that attention between groups of pixels sharing similar feature values can replace pixel-pair attention in semantic segmentation, cutting computational cost from quadratic to roughly linear while maintaining or improving…","keywords":["semantic segmentation","remote sensing imagery","threshold attention mechanism","self-attention","global context modeling","pyramid pooling","efficient attention","high-resolution urban scenes"],"falsifier":"Replace the threshold-level assignment in Eq. (7) with a random assignment of pixels to the same number $L$ of levels while keeping all other machinery fixed: if segmentation accuracy stays about the same, the value-based grouping itself is doing no work and the mechanism is just a cheap low-rank attention. The paper's own observation that the tree class is its weakest, attributed to thin branches and wide color variation, offers a second check: on classes with high intra-class color spread, TAM should systematically lose to pixel-pair attention, and a per-class accuracy comparison on those regions would expose the grouping failure.","tokens_in":16267,"feed_emoji":"🛰️","tokens_out":7290,"duration_ms":66208,"temperature":0.7,"pith_summary":"This paper tries to establish that a segmentation network can compute global attention between groups of pixels sharing similar feature values instead of between every pixel pair, and that this substitution both lowers cost and improves accuracy on high-resolution remote sensing images. The proposed threshold attention mechanism (TAM) quantizes each feature channel into $L$ value intervals, computes an $L\\times L$ attention matrix between intervals, then reattaches each pixel to its interval through a one-hot location code. Wrapped into two modules—AFEM for shallow features and TAPP for deep multi-scale features—TAM forms TANet, which the paper reports achieves top or near-top mean F1, overall accuracy, and mIoU on the Vaihingen and Potsdam benchmarks while using roughly one-twelfth the GFLOPs of standard self-attention. The result matters because global context in segmentation has until now come with quadratic pixel-pair cost, limiting use on very large imagery.","feed_headline":"Pixels grouped by value make attention linear, not quadratic","feed_subtitle":"It tops most metrics on two remote sensing benchmarks while using a fraction of the compute.","key_machinery":"The load-bearing object is the threshold attention mechanism, an attention layer whose tokens are per-channel value intervals rather than pixels. For input features $F\\in\\mathbb{R}^{C\\times N}$ with $N=H\\times W$, the mechanism computes per-channel thresholds that split each channel into $L$ equal-width intervals, obtaining $T\\in\\mathbb{R}^{L\\times C}$; it then applies query, key, and value projections and row-softmax to get an $L\\times L$ attention matrix $A$, so the cost depends on $L$ rather than on $N^2$. Pixel locations are preserved by the discretization map $P\\in\\mathbb{R}^{C\\times N}$, one-hot encoded and multiplied against the transposed attention matrix, which gives each pixel the attention weight of its bin. The same mechanism is reused in AFEM, paired with channel attention and a residual skip to enhance shallow features, and in TAPP, paired with dilated convolutions and global pooling to enrich deep multi-scale features.","core_discovery":"The central claim is that dependencies between spatially distant pixels can be modeled faithfully at the level of value bins: pixels whose feature vectors fall in the same per-channel interval belong to a 'threshold region,' and attention between these regions carries the global information the network needs. TAM performs this in three moves: uniform per-channel quantization to a threshold matrix $T\\in\\mathbb{R}^{L\\times C}$, linear projections and softmax to produce an $L\\times L$ attention matrix, and one-hot level assignment $P$ that scatters the region-level attention back onto individual pixels so the output has the same shape as the input. The paper reports that TANet built from AFEM and TAPP exceeds the listed comparison methods on most overall metrics, and that the two threshold-attention modules outweigh classical context and attention modules in the Vaihingen ablation.","pith_inferences":["The paper does not test TAM against alternative grouping rules, such as clustering by learned prototypes or per-class thresholds; a natural follow-up is whether equal-width value bins are the best partition or merely a cheap one.","Because the efficiency comparison excludes the backbone and the reported per-image inference time is not the fastest, the practical speed advantage over other full networks may be smaller than the GFLOPs table suggests.","The mechanism's grouping premise suggests a stress test on images with strong illumination gradients or sensor noise, where equal-width bins may split one object across many levels; adaptive or normalized binning would be a testable fix.","The same quantization-then-attention pattern could apply to other dense prediction tasks such as panoptic segmentation or depth estimation, where global region dependencies matter; the paper only demonstrates semantic segmentation."],"forward_implications":["Global attention for segmentation becomes affordable at high resolution: TAM's cost scales with the chosen number of value levels $L$ and linearly with pixel count, rather than quadratically.","The same modules can be dropped into other encoder-decoder segmentation networks because AFEM and TAPP preserve input and output shape and are reported to improve other standard backbones too.","The level count $L$ is a meaningful hyperparameter: optimal values differ between modules and datasets, so threshold granularity can be tuned per task.","Standard training add-ons such as auxiliary loss, online hard example mining, and test-time augmentation stack on top of TANet's gains, implying the region-level attention is complementary to loss- and inference-time improvements."],"supporting_citations":[{"why":"Defines the pixel-pair self-attention baseline whose quadratic cost and redundancy TAM is designed to replace.","marker":"[20]"},{"why":"Provides the ASPP context module that TAPP extends and serves as the DeepLabV3+ comparison in the benchmark tables.","marker":"[36]"},{"why":"Supplies the pyramid pooling context module used as an ablation baseline and as PPM in efficiency and benchmark comparisons.","marker":"[37]"},{"why":"Dual-attention baseline and the source of the Vaihingen and Potsdam training splits used in the experiments.","marker":"[35]"},{"why":"Recent remote-sensing hybrid attention network whose CAA&RSA module is a direct accuracy and efficiency comparison.","marker":"[19]"},{"why":"Multiattention network with linear-complexity kernel attention; supplies the efficiency comparison protocol and a benchmark rival.","marker":"[34]"},{"why":"External attention, an alternative efficient attention baseline included in the Vaihingen module ablation.","marker":"[30]"},{"why":"High-resolution boundary-constrained network used as a strong comparison in both benchmark tables.","marker":"[24]"}],"fun_headline_variants":["Value-binned attention makes segmentation linear, not quadratic","TANet: threshold attention cuts compute, tops remote sensing","Bin pixels by value, attend globally, segment faster","Threshold attention: O(n) compute, SOTA on Vaihingen and Potsdam"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that pixels falling in the same per-channel value interval are similar enough semantically that attention between intervals can replace attention between pixels; if binning destroys the correlations needed for accurate segmentation, the efficiency gain is bought at the cost of accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Value-binned attention makes segmentation linear, not quadratic","TANet: threshold attention cuts compute, tops remote sensing","Bin pixels by value, attend globally, segment faster","Threshold attention: O(n) compute, SOTA on Vaihingen and Potsdam"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1261,"prompt_tokens":923,"completion_tokens":338,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":265}},"tokens_in":539,"tokens_out":338,"duration_ms":4207,"temperature":1.0,"reasoning_tokens":265,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:29:44.071338+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the threshold-level assignment in Eq. (7) with a random assignment of pixels to the same number $L$ of levels while keeping all other machinery fixed: if segmentation accuracy stays about the same, the value-based grouping itself is doing no work and the mechanism is just a cheap low-rank attention. The paper's own observation that the tree class is its weakest, attributed to thin branches and wide color variation, offers a second check: on classes with high intra-class color spread, TAM should systematically lose to pixel-pair attention, and a per-class accuracy comparison on those regions would expose the grouping failure.","supporting_citations":[{"cited_title":"Dual attention network for scene segmentation,","cited_arxiv_id":null,"evidence_quote":"Dual-attention baseline and the source of the Vaihingen and Potsdam training splits used in the experiments."},{"cited_title":"Hybrid multiple attention network for semantic segmentation in aerial images,","cited_arxiv_id":null,"evidence_quote":"Recent remote-sensing hybrid attention network whose CAA&RSA module is a direct accuracy and efficiency comparison."},{"cited_title":"Multiattention network for semantic segmentation of fine-resolution remote sensing images,","cited_arxiv_id":null,"evidence_quote":"Multiattention network with linear-complexity kernel attention; supplies the efficiency comparison protocol and a benchmark rival."},{"cited_title":"Beyond self-attention: External attention using two linear layers for visual tasks,","cited_arxiv_id":null,"evidence_quote":"External attention, an alternative efficient attention baseline included in the Vaihingen module ablation."},{"cited_title":"High-resolution boundary-constrained and context- enhanced network for remote sensing image segmentation,","cited_arxiv_id":null,"evidence_quote":"High-resolution boundary-constrained network used as a strong comparison in both benchmark tables."}],"review_version":1}