{"id":"69497500-ed0c-4225-9ca5-ea587b20118e","arxiv_id":"2507.16251","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"HoliTracer is a framework that directly vectorizes geographic objects from very high resolution images over 10,000 pixels wide, using multi-scale attention and contour-sequence tracing.","lead":"A new computer vision system called HoliTracer extracts vector maps of buildings, water bodies, and roads directly from very large satellite images instead of cutting them into small patches. The authors report large gains over existing methods and release code and data, but the comparison may be favored by using a remote sensing pretrained model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'outperforms SOTA' claim is not yet supported because HoliTracer's encoder uses Skysense remote-sensing pretraining while all baselines use ImageNet backbones; this uncontrolled variable could explain much of the reported margin.","rationale":"The reader's weakest_assumption identifies the Skysense pretraining confound, and I agree it is the most load-bearing issue. The paper provides code and data, and the method is clearly described, but the headline empirical claim rests on Table 1, which does not isolate the contribution of the proposed components from the encoder's prior. The MCR algorithm also contains a minor index inconsistency (Algorithm 1 assumes M distinct matched vertices while the text allows P ≤ M), but this is a reproducibility issue rather than a threat to the central performance claim. I therefore recommend keeping the CONDITIONAL verdict, with the condition being a controlled pretraining comparison.","tokens_in":17202,"tokens_out":4320,"duration_ms":46938,"concrete_test":"Retrain the HoliTracer pipeline on WHU-building with the Swin-L encoder initialized from ImageNet-22K instead of Skysense, keeping all other hyperparameters and training settings identical, and report the same metrics as Table 1. If the IoU/CIoU/AP values drop toward the HiSup baseline such that the gap narrows by more than half, the reported superiority is largely an artifact of pretraining. As an additional cross-check, initialize HiSup with the same Skysense encoder and retrain; if it closes the gap, the comparison is unfair.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim in the abstract and Section 4.2 is that HoliTracer outperforms state-of-the-art methods on large-size RSI vectorization. This comparison is confounded by the encoder initialization. Section 3.1 states the encoder is Swin-L with Skysense weights, a remote sensing foundation model, whereas the compared methods (HiSup, DeepSnake, E2EC, FFL, UniVec, TS-MTA, LCF-ALE) use their default backbones pretrained on ImageNet. Table 5's ablation (HRNet+PST vs CAN+PST) controls for the PST module but not for pretraining, since HRNet is not Skysense-initialized. With 311.76M parameters versus 74.29M for HiSup (Table 7), HoliTracer's advantage in Table 1 (e.g., +5.81 IoU, +4.30 AP, +20.04 CIoU on WHU-building) may be substantially driven by these confounding factors rather than by CAN's local-to-global attention or MCR/PST. Until a controlled comparison is reported, the 'outperforms SOTA' claim should be treated as conditional.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HoliTracer, a framework for vectorizing geographic objects from large-size remote sensing imagery. It combines a Context Attention Net (CAN) that fuses multi-scale pyramid features for segmentation, a Mask Contour Reformer (MCR) that converts segmentation masks into polygon sequences aligned with ground-truth polygons, and a Polygon Sequence Tracer (PST) that refines vertices and classifies them. Experiments on WHU-building, GLH-water, and a newly introduced VHR-road dataset compare HoliTracer with existing vectorization and segmentation methods across vector, instance, and semantic metrics. The paper claims that HoliTracer is the first framework designed for holistic vector extraction from large-size RSI and that it outperforms state-of-the-art methods.","tokens_in":17481,"tokens_out":7923,"duration_ms":84667,"significance":"If the reported results are validated under controlled conditions, the paper would make a practically useful contribution to large-size remote sensing vectorization, a task where patch-based pipelines often lose context and produce fragmented outputs. The paper's strengths include a clear modular design, detailed ablations of PST and the multi-scale pyramid, a new large-size road dataset, and a public code/data release. However, the central empirical claim currently rests on an uncontrolled comparison: HoliTracer uses a Swin-L encoder initialized with remote-sensing foundation model weights (Skysense) while baselines use standard ImageNet-pretrained backbones and, in several cases, much smaller models. In addition, the MCR matching procedure contains an indexing inconsistency that affects the construction of training supervision. These issues are fixable and do not invalidate the overall approach, but they must be addressed before the reported performance advantage can be considered established.","major_comments":[{"comment":"The 'outperforms state-of-the-art' claim is confounded by encoder initialization and capacity. Section 3.1 states 'The encoder is Swin-L with Skysense [11] weights,' while the compared methods use their standard ImageNet-pretrained backbones (e.g., HiSup with HRNet, DeepSnake/E2EC with their default encoders). Table 7 reports 311.76M total parameters for HoliTracer versus 74.29M for HiSup. Table 5's ablation of HRNet+PST versus CAN+PST changes both the backbone and the pretraining, so it does not isolate the contribution of CAN. Please provide controlled comparisons with the same backbone and initialization protocol, such as Swin-L with ImageNet versus Swin-L with Skysense, and ideally the same training and inference tile sizes. Without such controls, the reported margins in Table 1 (e.g., CIoU 82.30 versus 62.26) may be substantially attributable to pretraining and model size rather than to the proposed components.","section":"Section 3.1 and Tables 1, 5, 7"},{"comment":"There is an indexing inconsistency in the MCR matching procedure. The text defines P as the number of unique matched reconstructed vertices, with indices i1 < ... < iP, and explicitly allows P <= M when multiple ground-truth vertices map to the same reconstructed point. However, the subsequent formulas use M in place of P: the text refers to i_{M+1} = i_1 and Eq. (4) writes |G'| = M + sum_{k=1}^M n_k = N. If P < M, this equality cannot hold because the reconstructed polygon has only P matched vertices and P edges between them. The construction of G' is load-bearing because it provides the point-wise supervision targets for PST. Please correct the indexing, define the correspondence between the P matched vertices and the original ground-truth vertices, and specify exactly how G' is formed in the collapsed-vertex case.","section":"Section 3.2 and Supplementary Algorithm 1, Eq. (4)"},{"comment":"The inference protocol for large-size images is underspecified, which is important because the paper's central claim is about holistic processing without patch stitching. The text says that a sliding window of uniform size slices the image pyramid, but it does not state the window size, stride, or overlap, nor does it explain how the per-window attention outputs are merged into a full-size segmentation mask. Table 7 states that HoliTracer 'processes large images directly, eliminating the need for patch-wise inference and subsequent stitching,' but the method description as written still appears to rely on sliding windows over the pyramid. Please clarify the exact training and inference tiling scheme, including how the pyramid patches are assembled, so that the 'no stitching' claim can be evaluated and reproduced.","section":"Sections 3.1 and 4.2, Table 7"}],"minor_comments":[{"comment":"The angle features are said to be 'converted to polar coordinate space,' but the conversion itself is not defined. Please specify the transformation (e.g., whether the angles are used as polar angles in an (r, theta) representation or simply concatenated as scalar features).","section":"Section 3.3"},{"comment":"The symbol d in QK^T / sqrt(d) is not defined in the text; it should be identified as the feature dimension.","section":"Equation (1)"},{"comment":"The image pyramid ablation reports only semantic metrics (IoU and F1). Since the paper's headline claims concern vector and instance metrics, reporting CIoU and AP for this ablation would strengthen the connection between CAN's context fusion and the vectorization improvements.","section":"Tables 4 and 9"},{"comment":"The label 'GroudTruth' in Figures 4 and 7 contains a typo; it should read 'GroundTruth.'","section":"Figures 4 and 7"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the Skysense initialization is valid and is the primary reason for major revision. The MCR indexing issue is also real and should be corrected. I do not see circularity in the empirical evaluation; the ablations are standard practice. The paper may be publishable after a controlled comparison is provided and the inference protocol is clarified, even if the performance gap narrows."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nTwo things to know about HoliTracer. The pipeline is a serious attempt at a real bottleneck: vectorizing remote sensing images over 10,000 pixels wide without the splitting artifacts of patch-based methods. And the paper's headline claim—that it outperforms state-of-the-art methods—is not yet supported because the comparison is confounded by encoder pretraining.\n\nWhat's genuinely new: the combination of a multi-scale image pyramid with a local-to-global attention segmentation net (CAN), followed by a contour reconstruction step (MCR) and a sequence-based vertex tracer (PST). That specific system for large-size RSI vectorization is new as far as I know. The authors also contribute a new road dataset (VHR-road) and release code. The ablations are a real strength: PST consistently improves vector metrics over Douglas-Peucker on the same segmentation output, the angle features and loss contribute, and Table 5 shows PST's benefit is not tied to a particular segmentation backbone. The engineering is solid and described in enough detail to reproduce.\n\nThe soft spot is the comparison. HoliTracer's encoder is Swin-L initialized with Skysense, a remote sensing foundation model; the baselines use standard ImageNet-pretrained backbones. That's a major uncontrolled variable. With 311M parameters versus 74M for HiSup, the reported gains (e.g., +20 CIoU, +5.8 IoU) could shrink substantially in a controlled comparison. The ablations don't fix this: HRNet+PST versus CAN+PST changes both the segmentation network and the pretraining. I'd want to see either the baselines re-run with Skysense backbones or HoliTracer trained from ImageNet before trusting the \"outperforms SOTA\" claim. There are also minor issues: a small inconsistency between Section 3.2 and Algorithm 1 in the MCR matching (the algorithm assumes a fixed number of unique vertices while the text allows duplicates), and no error bars or repeated runs.\n\nOverall, the paper is coherent and the design is well-motivated. The confound is fixable with extra experiments, and the dataset plus the PST module are valuable on their own. A serious editor should send this to peer review; it needs major revision but not desk rejection. Readers in large-image vectorization or GIS will get useful ideas and data from it, though they should treat the performance numbers as provisional.\n\nRegards.","headline":"A useful large-image vectorization pipeline whose headline SOTA numbers are not credible until the pretraining confound is controlled.","tokens_in":17980,"tokens_out":4605,"would_cite":false,"duration_ms":43361,"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":"HoliTracer claims a single segmentation-to-tracing pipeline turns 10,000×10,000-pixel remote sensing images directly into vector polygons for buildings, water, and roads.","keywords":["holistic vectorization","large-size remote sensing imagery","polygon extraction","context attention network","local-to-global attention","vector map generation","building water road extraction","polygon sequence tracing"],"falsifier":"Re-run the WHU-building comparison twice: once with HoliTracer's encoder initialized from ImageNet weights, and once with the leading baseline initialized from the same remote-sensing pretrained weights. If the reported gaps (CIoU 82.30 vs 62.26, AP 61.07 vs 56.77) collapse or reverse under matched initialization, the central outperformance claim is not attributable to the framework.","tokens_in":16997,"feed_emoji":"🗺️","tokens_out":8075,"duration_ms":73799,"temperature":0.7,"pith_summary":"The paper sets out to solve a practical bottleneck: remote sensing images now routinely exceed 10,000 × 10,000 pixels, but existing vectorization methods are built for 512 × 512 patches, so they lose surrounding context and stitch together fragmented polygons. HoliTracer is proposed as the first framework that takes the whole large image as input and outputs complete vector polygons for multiple object classes in one go. The authors argue that a segmentation stage with a local-to-global attention mechanism preserves the context that patch methods drop, and that a two-part polygon reformer then produces clean, ground-truth-aligned vectors. On large-size datasets of buildings, water bodies, and roads, they report that HoliTracer beats current state-of-the-art methods on vector, instance, and semantic metrics. The practical stake is that vector maps for navigation and planning could be produced directly from very large imagery without patch-boundary artifacts.","feed_headline":"One pipeline vectorizes whole 10,000-pixel satellite images","feed_subtitle":"Buildings, water, and roads come out as complete polygons instead of fragmented patch-boundary pieces, the paper reports.","key_machinery":"The framework rests on three modules. The Context Attention Net slices a multi-scale image pyramid with a shared Swin-L encoder and fuses bottom-layer features with coarser-layer features via attention, so each local patch receives global surroundings. The Mask Contour Reformer reconstructs polygons by simplifying mask contours with the Douglas-Peucker algorithm, re-sampling points at a fixed interpolation distance, and using bidirectional nearest-point matching to generate supervised training pairs aligned with ground-truth polygons. The Polygon Sequence Tracer refines those reconstructed polygons through iterative offset regression and vertex classification, with angle features and an angle-penalty loss that make small-angle corner points more likely to be selected as vertices.","core_discovery":"The core claim is that holistic vectorization of large-size remote sensing imagery is achievable with a segmentation-first, trace-second pipeline. HoliTracer first segments the entire large image with the Context Attention Net, which uses an image pyramid and a local-to-global attention mechanism to fuse fine local detail with coarse global context. The resulting mask is converted to polygon contours by the Mask Contour Reformer, which simplifies raw contours, interpolates points at a fixed spacing, and during training aligns them one-to-one with ground-truth polygons through bidirectional matching. The Polygon Sequence Tracer then treats the polygon as a sequence, regresses per-point offsets in several iterative rounds, and classifies each point as a vertex using angle features and an angle-penalty loss. The paper reports that this combination yields stronger completeness and geometric fidelity than patch-based comparison methods on buildings, water bodies, and roads.","pith_inferences":["A controlled experiment with matched encoder initialization would be needed to separate the contribution of the HoliTracer modules from that of the remote-sensing-pretrained backbone; the paper does not report such an ablation, so the size of the true architectural gain is untested.","The same Context Attention Net plus Mask Contour Reformer plus Polygon Sequence Tracer structure could plausibly be applied to other polygonizable land-cover classes, such as agricultural fields or parking lots, by adjusting only the pyramid scales and interpolation distance.","The pyramid-based full-image context suggests a path toward streaming inference: tiling only the computing, not the perception, so arbitrarily large imagery keeps global context without a hard memory ceiling."],"forward_implications":["Large-size RSI can be vectorized without patch-boundary fragmentation, so building, water, and road outlines keep their geometric integrity across the full scene.","A single multi-class pipeline can output unified vector maps rather than separate single-object extractors, simplifying downstream mapping workflows.","Local-to-global attention helps disambiguate objects that look similar at patch scale, such as buildings versus parking lots.","The angle-aware vertex predictor makes polygon complexity track ground-truth complexity, reducing both over-simplified and over-complicated outlines.","Because inference runs directly on large images, the method scales to very large scenes (the paper demonstrates up to 40,000 × 50,000 pixels) without stitching overhead."],"supporting_citations":[{"why":"Supplies the remote-sensing pretraining that initializes the Swin-L encoder and is not applied to the compared baselines.","marker":"[11]"},{"why":"Defines the Swin Transformer backbone used by the Context Attention Net.","marker":"[26]"},{"why":"Provides the UperNet segmentation head used to decode the fused multi-scale features.","marker":"[36]"},{"why":"Supplies the TC89-KCOS algorithm that extracts initial polygon contours from segmentation masks.","marker":"[30]"},{"why":"Supplies the Douglas-Peucker algorithm used to simplify contours during Mask Contour Reformer reconstruction.","marker":"[8]"},{"why":"Inspires the iterative offset-regression refinement used by the Polygon Sequence Tracer.","marker":"[27]"},{"why":"HiSup is the strongest patch-based building vectorization baseline HoliTracer compares against.","marker":"[37]"},{"why":"FFL is a segmentation-based vectorization baseline that must be outperformed on the large-size datasets.","marker":"[10]"},{"why":"UniVec is the graph-based universal vector extraction baseline included in the main comparison.","marker":"[41]"}],"fun_headline_variants":["HoliTracer vectorizes entire large satellite images in one pass","First framework to vectorize entire large remote sensing images","Whole-scene vectorization for large remote sensing imagery","One pipeline turns large satellite images into clean vector maps","From patchwork fragments to complete polygons in large images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is comparison fairness: HoliTracer's encoder is initialized with remote-sensing pretrained weights while the baselines use standard pretrained backbones, so the reported performance gap could come from that initialization rather than from the proposed modules.","fun_headline_variants_meta":{"raw":{"variants":["HoliTracer vectorizes entire large satellite images in one pass","First framework to vectorize entire large remote sensing images","Whole-scene vectorization for large remote sensing imagery","One pipeline turns large satellite images into clean vector maps","From patchwork fragments to complete polygons in large images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001017,"raw_usage":{"total_tokens":4277,"prompt_tokens":916,"completion_tokens":3361,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":3282}},"tokens_in":532,"tokens_out":3361,"duration_ms":27115,"temperature":1.0,"reasoning_tokens":3282,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:14:20.402171+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the WHU-building comparison twice: once with HoliTracer's encoder initialized from ImageNet weights, and once with the leading baseline initialized from the same remote-sensing pretrained weights. If the reported gaps (CIoU 82.30 vs 62.26, AP 61.07 vs 56.77) collapse or reverse under matched initialization, the central outperformance claim is not attributable to the framework.","supporting_citations":[{"cited_title":"Skysense: A multi-modal remote sens- ing foundation model towards universal interpretation for earth observation imagery","cited_arxiv_id":null,"evidence_quote":"Supplies the remote-sensing pretraining that initializes the Swin-L encoder and is not applied to the compared baselines."},{"cited_title":"Unified perceptual parsing for scene understand- ing","cited_arxiv_id":null,"evidence_quote":"Provides the UperNet segmentation head used to decode the fused multi-scale features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the TC89-KCOS algorithm that extracts initial polygon contours from segmentation masks."},{"cited_title":"Algorithms for the reduction of the number of points required to represent a dig- itized line or its caricature","cited_arxiv_id":null,"evidence_quote":"Supplies the Douglas-Peucker algorithm used to simplify contours during Mask Contour Reformer reconstruction."},{"cited_title":"Deep snake for real-time instance seg- mentation","cited_arxiv_id":null,"evidence_quote":"Inspires the iterative offset-regression refinement used by the Polygon Sequence Tracer."},{"cited_title":"Hisup: Accurate polygonal mapping of buildings in satellite im- agery with hierarchical supervision","cited_arxiv_id":null,"evidence_quote":"HiSup is the strongest patch-based building vectorization baseline HoliTracer compares against."},{"cited_title":"Polygonal building extraction by frame field learning","cited_arxiv_id":null,"evidence_quote":"FFL is a segmentation-based vectorization baseline that must be outperformed on the large-size datasets."},{"cited_title":"Univecmapper: A universal model for thematic and multi-class vector graph extraction","cited_arxiv_id":null,"evidence_quote":"UniVec is the graph-based universal vector extraction baseline included in the main comparison."}],"review_version":1}