{"id":"bb87de88-d4f3-47bc-90b3-1a7e76e66c36","arxiv_id":"2505.23597","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A tree crown segmentation model using a trainable Log-Gabor convolutional layer, mixed pooling, and averaged dilated convolutions reports mIoU gains over ResUNet and other baselines on three aerial datasets.","lead":"PerceptiveNet is a modified ResUNet that swaps its first layer for a trainable Log-Gabor filter, adds mixed pooling and averaged dilated convolutions, and reports higher mIoU than several baselines on three aerial datasets. The paper also presents a CNN-Transformer hybrid, PerceptiveNeTr, which underperforms the pure CNN version.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on unreplicated, statistically untested mIoU differences, and the split protocol for Landcover.AI and UA Vid is not stated, so test-set leakage or run-to-run noise could explain the reported margins.","rationale":"The reader's weakest_assumption identifies exactly the evaluation-reliability problem: single-run mIoU differences without confidence intervals and unspecified patch-level splitting. My independent read of the paper reaches the same conclusion, with the additional observation that the datasets themselves create concrete leakage mechanisms. TreeCrown is a multi-month acquisition of the same forest, Landcover.AI is subsampled from large orthophotos, and UA Vid is produced by cropping original high-resolution frames; in all three cases a random image/crop split is a realistic default that would let near-duplicate content appear in both training and test sets. Because the paper does not describe the split unit, the reported 2–5 point improvements cannot be trusted as evidence of true generalisation. This is not a disagreement with the architectural idea or the theoretical motivation for Log-Gabor layers; it is a failure of the evidence to support the specific quantitative claim. The reader's CONDITIONAL verdict is appropriate: the paper should be accepted only if the authors add repeated-run statistics and explicit source-level splits, or temper the claims to match the evidence. No additional concern beyond the reader's is needed to support this verdict, so I mark agreement as 'agree' and keep the verdict unchanged.","tokens_in":12481,"tokens_out":3939,"duration_ms":40704,"concrete_test":"Retrain every model in Tables 1–4 with at least 5 fixed seeds, reporting mean±std mIoU and a paired significance test (e.g., corrected resampled t-test or Wilcoxon signed-rank) between PerceptiveNet and the best baseline. In parallel, re-run the Landcover.AI and UA Vid experiments with splits performed at the orthophoto/video-frame level rather than at the crop level, and for TreeCrown split by acquisition date or explicitly ensure no crown appears in both train and test. If PerceptiveNet's margin over the best baseline is not consistently positive and significant under these protocols, the SOTA claim in Section 3.4 should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that PerceptiveNet outperforms all models on TreeCrown, Landcover.AI, and UA Vid is supported only by single-run mIoU values. Sections 3.2 and 3.4 report differences of roughly 2–5 mIoU points over the strongest baselines (e.g., 48.1 vs. 45.9 for MACUNet on TreeCrown; 81.6 vs. 78.6 for DeepLabV3+ on Landcover.AI), yet no seeds, repeated runs, confidence intervals, or significance tests are provided. Training from scratch with Adam and no stated learning-rate schedule is known to exhibit run-to-run variance on the order of several mIoU points, so the headline improvements may not be reproducible. The paper also never specifies whether Landcover.AI and UA Vid patches are split by source orthophoto/video or by random crop. Landcover.AI was sub-sampled from 41 orthophotos into 21,924 crops; if crops from the same orthophoto appear in both train and test, spatially correlated pixels inflate mIoU. UA Vid images were padded and cropped into eight 256×256 patches each; if patches from the same original frame are not kept in the same fold, the test set overlaps the training distribution at the frame level. The TreeCrown dataset likewise contains repeated monthly acquisitions of the same stands, so a random image-level split can put the same crowns in both train and test. Without a clear split-by-source protocol and error bars, the claimed state-of-the-art performance is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PerceptiveNet, a U-Net-style segmentation architecture whose initial convolution is parameterized by trainable Log-Gabor filters, and which incorporates residual blocks with mixed max-average pooling and averaged dilated convolutions. It also introduces PerceptiveNeTr, a hybrid CNN-Transformer variant using PerceptiveNet as the backbone. The authors compare against ResUNet, U-Net, MACUNet, DeepLabV3+, UNETR 2D, and SwinUNet on three aerial datasets (TreeCrown, Landcover.AI, UA Vid), reporting mIoU improvements of roughly 2.5 to 10.5 points over the strongest baselines, with an ablation study that decomposes the contributions of the proposed components. The central claim is that PerceptiveNet outperforms all compared models on all three datasets and that the Log-Gabor layer provides a consistent gain over standard and Gabor-based convolutions.","tokens_in":13014,"tokens_out":5127,"duration_ms":49531,"significance":"If the reported results are reproducible, the work offers a simple and well-motivated architectural modification—trainable Log-Gabor filters—that improves segmentation on aerial forest and urban scenes, with a clear theoretical rationale based on zero DC components and more uniform frequency coverage. The multi-dataset evaluation and the ablation isolating pooling, dilated convolution, and Log-Gabor effects are useful contributions. However, the empirical evidence is currently thin: all numbers come from single training runs, no confidence intervals or significance tests are provided, the train/test split protocol for the two benchmark datasets is not described, and key hyperparameters are hand-picked. Because the headline margins are only a few mIoU points, the claimed state-of-the-art status is not yet established. The paper does not provide code or trained models, which further limits reproducibility.","major_comments":[{"comment":"The central claim that PerceptiveNet outperforms all compared models rests entirely on single-run mIoU values. For example, the tree-crown margin over MACUNet is 2.2 mIoU points (48.1 vs. 45.9), the Landcover.AI margin over DeepLabV3+ is 3.0 points (81.6 vs. 78.6), and the UA Vid margin over MACUNet is 2.5 points (68.6 vs. 66.1). Training from scratch with Adam and no stated learning-rate schedule can produce run-to-run variation of several mIoU points. Please report mean and standard deviation over at least three to five seeds, and provide a significance test or an explicit statement of variance; otherwise the reported ranking may not be reproducible.","section":"§3.4, Tables 3 and 4"},{"comment":"The paper does not state whether the Landcover.AI and UA Vid splits are performed by source orthophoto/video or by random crop. Landcover.AI is sub-sampled from 33+8 large orthophotos into 21,924 crops; if crops from the same orthophoto appear in both training and test sets, spatially correlated pixels can inflate mIoU. Similarly, UA Vid is cropped into eight patches per original frame, and a random patch-level split can place patches from the same frame in both training and test. TreeCrown has repeated monthly acquisitions of the same stands, so a random image-level split may leak the same crowns across the boundary. Please specify the split protocol explicitly and, ideally, re-run with source-image/frame-level splits to rule out leakage.","section":"§2.2, Data"},{"comment":"The mixing ratio α=0.8, the dilation rates 1, 3, 6, 9, the four transformer layers, and the 130 training epochs are presented as fixed choices without sensitivity analysis. Since these values were apparently selected using the validation sets before reporting test scores, and since only one configuration is evaluated, the reported differences may reflect selection bias rather than a robust architectural advantage. Please provide an ablation over α and the dilation-rate setting, or at least report results for multiple seeds under the chosen configuration, to show that the conclusion is not an artifact of a single hand-picked configuration.","section":"§2.1, Model Training and hyperparameters"},{"comment":"The hybrid CNN-Transformer model is described only at a high level: patch embedding, LayerNorm, MSA, and MLP are mentioned, but the patch size, embedding dimension, number of heads, MLP hidden dimension, and positional encoding are not specified. Without these details, the transformer comparison in Table 4 cannot be reproduced, and the reported gains of PerceptiveNeTr over ViTResUNet cannot be attributed to the backbone rather than to differences in the transformer configuration. Please add a complete specification of the transformer hyperparameters.","section":"§2.1, PerceptiveNeTr and Figure 4"}],"minor_comments":[{"comment":"There is a notation inconsistency in the Log-Gabor layer: the text lists f as a learnable parameter, but Eq. (4) and Eq. (5) use only f0 and σ in the radial and cosine terms, and no equation contains f. Please clarify whether f is a separate learnable parameter or a typo, and define all symbols in one place.","section":"§2.1, Eqs. (4)–(6)"},{"comment":"The dataset name 'UA Vid' is written inconsistently; the original source is 'UAVid'. Please use the official name for consistency and to aid readers in locating the dataset.","section":"§2.2 and throughout"},{"comment":"Calling ViTResUNet a 'Proposed Hybrid CNN-Transformer' is misleading because ViTResUNet is the ResUNet-backbone baseline used to isolate the effect of the PerceptiveNet backbone. Please rephrase to avoid implying that ViTResUNet is a contribution.","section":"Table 4 and §3.4.2"},{"comment":"The qualitative claims about boundary quality and dead-tree segmentation are plausible but not quantified. Consider reporting a boundary-oriented metric (e.g., boundary IoU or contour F1) to support these visual observations.","section":"§3.5 and Figure 6"},{"comment":"Some references are incomplete; for example, [25] and [26] list only the title and venue without author names. Please complete the bibliography.","section":"References"},{"comment":"The learning rate is not reported, and it is unclear whether any learning-rate schedule is used. Since the paper notes that 130 epochs were chosen based on validation improvement, please state the exact training protocol, optimizer settings, and the criterion for stopping.","section":"§2.1, Model Training"}],"recommendation":"major_revision","confidential_remarks":"The idea is in scope for a computer-vision journal and the proposed modification is simple enough to be worth pursuing. The main obstacles are experimental: single-run results, an unspecified split protocol for Landcover.AI and UA Vid, and hand-picked hyperparameters without sensitivity checks. The split-by-source issue is the most serious, as it directly affects the validity of the generalization claim. I would advise requiring source-level splits and at least three seed runs, plus code or a detailed implementation specification, before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a solid incremental architecture paper whose headline SOTA claim is not yet supported by the evidence. The genuinely new piece is a trainable Log-Gabor parameterized first convolution layer combined with mixed pooling and averaged dilated residual blocks. The theory is standard and applied correctly, and the paper gives enough detail to re-implement. What it does well is a clean ablation isolating the three components across three datasets, plus CAMs and qualitative figures. The citation pattern is fine: GaborNet, Gabor Convolutional Networks, and Log-Gabor wavelets are all cited, and the self-citation to the author's earlier domain-shift paper is relevant.\n\nThe soft spots are real and load-bearing. Every reported number is from a single run. No seeds, no repeated runs, no confidence intervals, no significance tests. The margins over strong baselines are 2–5 mIoU points on TreeCrown and 1.5–3 points on the other datasets. With Adam training from scratch, run-to-run variance is commonly on that order, so the SOTA claim is fragile. Worse, the split protocol is underspecified. Landcover.AI is subsampled from 41 orthophotos into 21,924 crops; if random crops from the same orthophoto appear in both train and test, pixel-level correlation can inflate mIoU. UA Vid is cropped into eight patches per frame; if patches from the same frame are split across train and test, the test set overlaps the training distribution. TreeCrown has repeated monthly acquisitions of the same stands, so an image-level random split can leak. The paper does not say whether splits were done by source image or by crop. That is a straightforward fix, but it has to be done before the numbers mean anything.\n\nMinor points: α=0.8 and the dilation rates (1,3,6,9) appear hand-picked with no sensitivity analysis, and no code is released. For an empirical architecture claim, repeated runs with seeds and a lookup by source image are the price of admission. None of these are fatal to the idea itself, which is plausible, but they are fatal to the specific claim of outperforming SOTA on these benchmarks.\n\nWho is this for? People working on remote sensing segmentation, especially tree crown delineation, get a plausible new component and a useful baseline comparison. It is not a breakthrough, but it is not a toy either. The paper deserves a serious referee rather than a desk reject; the referee should ask for reruns with seeds, a split-by-source protocol, confidence intervals, and preferably a small sensitivity study on the mixing ratio and dilation rates before accepting the SOTA claim.","headline":"A plausible Log-Gabor backbone with a clean ablation, but the SOTA claim rests on single-run mIoU gaps and an underspecified split protocol.","tokens_in":13309,"tokens_out":2463,"would_cite":false,"duration_ms":25736,"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":"PerceptiveNet, a segmentation network whose first layer uses trainable Log-Gabor filters, reports the highest mean intersection-over-union scores on three aerial datasets, including a 48.1% mIoU on TreeCrown.","keywords":["tree crown segmentation","Log-Gabor filters","semantic segmentation","aerial imagery","deep learning","UAV remote sensing","CNN","transformer hybrid"],"falsifier":"Run each compared model at least five times with different random seeds and the same data split, then check whether PerceptiveNet's mIoU advantage over the closest baseline (e.g., MACUNet on TreeCrown) exceeds the standard deviation of the runs; additionally, re-split the data by orthophoto so that no patches from the same orthophoto appear in both train and test, and see whether the reported margins survive.","tokens_in":12310,"feed_emoji":"🌳","tokens_out":5488,"duration_ms":44372,"temperature":0.7,"pith_summary":"The paper argues that a convolutional layer parameterised by trainable Log-Gabor functions, inserted as the first layer of a ResUNet-style encoder-decoder, gives the network a more balanced frequency coverage and better shape-based features, and that combining this layer with mixed max-average pooling and averaged dilated convolutions yields a backbone that segments tree crowns in dense aerial forest imagery more accurately than standard U-Net variants and transformer hybrids. On the TreeCrown dataset the complete model, PerceptiveNet, reaches 48.1% mean IoU, which the paper reports as a 10.5 percentage-point improvement over its ResUNet baseline and above the other compared CNN models; the same backbone also posts the highest mIoU on Landcover.AI (81.6%) and UA Vid (68.6%). The paper further claims that the backbone transfers to a hybrid CNN-Transformer model, PerceptiveNeTr, which outperforms pure-transformer comparators. A sympathetic reader would care because tree-crown segmentation is the basis for forest health, biodiversity, and carbon accounting, and the proposed layer is a small architectural change that could be adopted wherever aerial imagery suffers from shadows, scale variation, and weak spectral differences.","feed_headline":"PerceptiveNet tops tree crown segmentation with 48.1% mIoU","feed_subtitle":"Trainable Log-Gabor filters plus a wider receptive field lift accuracy on aerial forest imagery and two more datasets.","key_machinery":"The load-bearing object is the Log-Gabor-parameterised convolutional layer: a 2D filter $g(x,y) = g_r(r) \\cdot g_\\theta(\\theta) \\cdot \\cos(2\\pi f_0 r + \\psi) / (2\\pi \\sigma^2)$, with radial and angular Gaussian envelopes in the log-frequency domain, whose parameters ($f, \\theta, \\sigma, \\psi, f_0, \\theta_0$) are initialised and then learned by backpropagation. Its stated theoretical advantages are a zero DC component, better orthogonality across scales, uniform Fourier-domain coverage, and better spatial localisation compared with ordinary Gabor filters. The layer is combined with two supporting mechanisms: a mixture of maximum and average pooling with mixing ratio $\\alpha = 0.8$ to widen the receptive field while halving resolution, and an averaged dilated convolutional block that sums the localisation maps from dilation rates 1, 3, 6, 9 to avoid spurious high-rate responses. Together these form the PerceptiveNet residual encoder; the decoder mirrors the residual block with the dilated layer. The hybrid PerceptiveNeTr appends four transformer layers (patch embedding, LayerNorm, multi-head self-attention, MLP) after the PerceptiveNet encoder.","core_discovery":"The central claim is that replacing the initial convolutional layer of a ResUNet with a layer whose filter weights are generated by a Log-Gabor function with trainable parameters—frequency, orientation, bandwidth, phase—and pairing that with residual blocks that use a mixture of maximum and average pooling and an averaged set of dilated convolutions (rates 1, 3, 6, 9) produces a segmentation model that captures both fine spatial detail and broad context. The paper reports PerceptiveNet achieving the highest mIoU of all compared models on all three datasets: 48.1% on TreeCrown, 81.6% on Landcover.AI, and 68.6% on UA Vid, with ablation results attributing gains to the Log-Gabor layer, the pooling mixture, and the dilated block in combination. It also reports that the same backbone inside a CNN-Transformer hybrid, PerceptiveNeTr, reaches 42.0%, 75.3%, and 64.3% mIoU on the three datasets, exceeding pure-transformer UNETR 2D and SwinUNet, and that Class Activation Maps show more focused activation on tree regions than the ResUNet baseline.","pith_inferences":["Editorial inference: if the Log-Gabor first layer is the true source of the gain, it should transfer to other encoder-decoder backbones (e.g., DeepLabV3+) when swapped into the first convolution; the paper does not test this, but it is a cheap experiment.","Editorial inference: the zero-DC property of Log-Gabor filters suggests the layer acts as a learned band-pass front-end; one could test this directly by visualising the learned filters' frequency responses after training to see whether they spread across scales as the theory assumes.","Editorial inference: the dataset's split method (by orthophoto vs by random crop) is not reported; if patches from the same orthophoto appear in both train and test, the reported mIoU could be inflated by spatial autocorrelation, so the strongest follow-up would evaluate on held-out orthophotos.","Editorial inference: because the paper reports only single-run results, the margins over the nearest baselines may not be stable; a prudent next step is to rerun each model with several seeds and report mean and standard deviation."],"forward_implications":["If the reported scores hold, PerceptiveNet would set the highest mIoU on TreeCrown, Landcover.AI, and UA Vid among the models compared, making the Log-Gabor first layer a drop-in alternative for aerial segmentation pipelines.","The ablation suggests the three components are complementary: removing any one lowers mIoU, so practitioners should keep all three rather than choosing one.","The backbone's success inside PerceptiveNeTr implies the benefit of Log-Gabor features is not tied to a pure-CNN decoder, but persists when global context is added by transformers.","The qualitative results on dead and overlapping trees suggest the model is better at high-frequency boundary detail, which is exactly where dense-forest canopies fail."],"supporting_citations":[{"why":"Supplies the TreeCrown dataset of 1,360 UAV images with 14 tree species classes, the primary benchmark.","marker":"[37]"},{"why":"Provides the Landcover.AI aerial dataset used to test generalisation to buildings, forests, water, and roads.","marker":"[38]"},{"why":"Provides the UA Vid urban UAV semantic segmentation dataset with eight class labels.","marker":"[39]"},{"why":"Defines the ResUNet architecture that PerceptiveNet extends and uses as the primary baseline.","marker":"[31]"},{"why":"Defines the U-Net architecture, a key convolutional baseline in the comparison tables.","marker":"[7]"},{"why":"Provides MACUNet, the closest CNN baseline on TreeCrown that PerceptiveNet claims to outperform.","marker":"[41]"},{"why":"Provides DeepLabV3+, a state-of-the-art semantic segmentation baseline with atrous separable convolution.","marker":"[42]"},{"why":"Establishes the self-invertible 2D Log-Gabor wavelet theory that the proposed convolutional layer is built on.","marker":"[28]"},{"why":"Introduces the averaged dilated convolution approach used in the residual blocks to enlarge the receptive field.","marker":"[35]"},{"why":"Defines the Vision Transformer whose encoder blocks are reused in the hybrid PerceptiveNeTr model.","marker":"[36]"}],"fun_headline_variants":["PerceptiveNet: Log-Gabor filters top tree crown mIoU","Trainable Log-Gabor layer lifts tree segmentation to 48.1%","PerceptiveNet: wider receptive field and Log-Gabor win","Log-Gabor convolution boosts tree crown segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation treats single-run mIoU differences as meaningful; the load-bearing premise is that a gap of a few points between models reflects real architectural superiority rather than run-to-run noise, and that the test patches do not share scenes with training patches.","fun_headline_variants_meta":{"raw":{"variants":["PerceptiveNet: Log-Gabor filters top tree crown mIoU","Trainable Log-Gabor layer lifts tree segmentation to 48.1%","PerceptiveNet: wider receptive field and Log-Gabor win","Log-Gabor convolution boosts tree crown segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000517,"raw_usage":{"total_tokens":2538,"prompt_tokens":1008,"completion_tokens":1530,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":1458}},"tokens_in":624,"tokens_out":1530,"duration_ms":10146,"temperature":1.0,"reasoning_tokens":1458,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:41:26.509899+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run each compared model at least five times with different random seeds and the same data split, then check whether PerceptiveNet's mIoU advantage over the closest baseline (e.g., MACUNet on TreeCrown) exceeds the standard deviation of the runs; additionally, re-split the data by orthophoto so that no patches from the same orthophoto appear in both train and test, and see whether the reported margins survive.","supporting_citations":[{"cited_title":"Influence of temperate forest autumn leaf phenology on segmentation of tree species from uav imagery using deep learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the TreeCrown dataset of 1,360 UAV images with 14 tree species classes, the primary benchmark."},{"cited_title":"Landcover.ai: Dataset for automatic mapping of buildings, woodlands, water and roads from aerial imagery,","cited_arxiv_id":null,"evidence_quote":"Provides the Landcover.AI aerial dataset used to test generalisation to buildings, forests, water, and roads."},{"cited_title":"Uavid: A semantic segmentation dataset for uav im- agery,","cited_arxiv_id":null,"evidence_quote":"Provides the UA Vid urban UAV semantic segmentation dataset with eight class labels."},{"cited_title":"Road extraction by deep residual u-net,","cited_arxiv_id":null,"evidence_quote":"Defines the ResUNet architecture that PerceptiveNet extends and uses as the primary baseline."},{"cited_title":"U-net: Con- volutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Defines the U-Net architecture, a key convolutional baseline in the comparison tables."},{"cited_title":"Multiattention network for semantic seg- mentation of fine-resolution remote sensing images,","cited_arxiv_id":null,"evidence_quote":"Provides MACUNet, the closest CNN baseline on TreeCrown that PerceptiveNet claims to outperform."},{"cited_title":"Encoder-decoder with atrous separable convolution for se- mantic image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides DeepLabV3+, a state-of-the-art semantic segmentation baseline with atrous separable convolution."},{"cited_title":"Self-invertible 2d log-gabor wavelets,","cited_arxiv_id":null,"evidence_quote":"Establishes the self-invertible 2D Log-Gabor wavelet theory that the proposed convolutional layer is built on."},{"cited_title":"Re- visiting dilated convolution: A simple approach for weakly- and semi-supervised semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Introduces the averaged dilated convolution approach used in the residual blocks to enlarge the receptive field."},{"cited_title":"An im- age is worth 16x16 words: Transformers for image recogni- tion at scale,","cited_arxiv_id":null,"evidence_quote":"Defines the Vision Transformer whose encoder blocks are reused in the hybrid PerceptiveNeTr model."}],"review_version":1}