{"id":"1a9849cd-e593-4ef4-b175-96db4269db22","arxiv_id":"2507.02307","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Flow-CDNet jointly trains an optical flow branch and a change detection branch to detect both slow displacements and fast appearance/disappearance changes in bitemporal images, reporting FEPE 0.869 on a self-built synthetic dataset.","lead":"An engineering paper proposes Flow-CDNet, a two-branch network that combines optical flow estimation with binary change detection so one model can spot both gradual displacement and abrupt appearance of objects in image pairs taken at different times. The authors also build a synthetic training dataset, a combined loss, and a new evaluation metric, and report gains over single-task baselines on that dataset plus a 20-pair real-world dam monitoring test.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world validation cannot substantiate the slow-change claim: unified GT masks and per-scene thresholds make the reported F1=0.8126 uninformative about whether the flow branch detects gradual deformation.","rationale":"The reader's weakest assumption was that the synthetic Flow-Change dataset is a faithful training/evaluation proxy for real slow and fast change scenarios. My concern is closely related but more pointed: even granting the synthetic dataset, the real-world evaluation in Section 4.3 cannot verify the slow-change claim because its ground truth is unified across change types and thresholds are selected per scene. This is the single most load-bearing weakness because the paper's practical value rests on detecting gradual deformation in monitoring scenarios, and the only real-world quantitative result is structurally incapable of demonstrating that. The paper does have independent supporting structure: the ablation in Table 2 is internally coherent, the synthetic results are plausible, and the efficiency analysis is concrete. However, the real-world evidence gap, combined with the internal inconsistencies noted by the reader (backbone identified as SpyNetC2 in Section 4.2 but RAFT in Sections 3.1.2 and 4.5.1; loss weight symbol mismatch; Eq. 9 missing the squared norm), keeps the central claims conditional rather than established. The reader's CONDITIONAL verdict already captures this appropriately, so I do not propose a verdict change.","tokens_in":15874,"tokens_out":3223,"duration_ms":41699,"concrete_test":"On the 20 real dam pairs, have independent annotators create separate GT masks for slow changes (deformation/displacement) and fast changes (appearance/disappearance). Run Flow-CDNet with a single fixed threshold, or with a pre-specified threshold-selection rule applied identically across scenes, and compute per-type recall and F1 for each branch separately. Compare these per-type scores against CD-only and flow-only baselines. If the per-type F1 for slow changes is not clearly above the CD-only baseline, the simultaneous-detection and generalization claims should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Flow-CDNet simultaneously detects slow and fast changes and generalizes to real dam imagery. Section 4.3 is the only evidence for real-world generalization, but its evaluation cannot support the slow-change half of the claim. The paper states that all observable change regions are annotated without categorizing them as abrupt or gradual, and that detections from either branch are considered valid if they localize these unified regions. The reported F1=0.8126 is obtained 'after selecting appropriate change thresholds for different scenes,' with no protocol for threshold selection, no held-out threshold tuning, no error bars, and no baseline comparison. Consequently, the metric cannot distinguish fast from slow detections, cannot attribute performance to the optical flow branch, and cannot demonstrate that gradual deformation is detected. Moreover, no real-world optical flow ground truth exists, so the 'usable flow field' component of the central claim is entirely unverified. If the unified GT is dominated by fast appearance/disappearance events, a change-detection-only model could match or exceed this F1, making the claimed mutual enhancement and real-world generalization an artifact of evaluation design rather than evidence of capability.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Flow-CDNet, a dual-branch network for detecting both 'slow' changes (object displacement) and 'fast' changes (appearance/disappearance) in bitemporal images. An optical flow branch (based on RAFT) and a binary change detection branch (based on ResNet/PSP) are trained jointly with an L2 loss on masked flow regions and a Tversky loss on change masks. To support training and evaluation, the authors construct a synthetic dataset Flow-Change from FlyingChairs and PASCAL VOC, introduce a composite loss, and propose a new metric FEPE. Quantitative results on Flow-Change show that the full model outperforms single-task baselines and the authors' own variants, and an ablation indicates mutual improvement between the two branches. A small real-world dataset of 20 dam-image pairs is used to demonstrate generalization, with a reported F1 of 0.8126.","tokens_in":15990,"tokens_out":10188,"duration_ms":104144,"significance":"The idea of jointly estimating optical flow and change detection is relevant for monitoring applications such as dam and slope surveillance, and the internal ablation (F1 rising from 0.753 to 0.892, mEPE falling from 2.409 to 1.027 when both branches are used) is a useful empirical observation. The new dataset and metric, if released and properly validated, could benefit the change-detection community. However, the evaluation is almost entirely in-house: the synthetic benchmark is constructed by the authors, the comparison set consists of single-task baselines and the authors' own variants, and the real-world validation uses a small sample with unified labels that cannot attribute detections to the flow branch. As a result, the central claim of robust simultaneous detection of slow and fast changes in real scenes is not yet substantiated.","major_comments":[{"comment":"The real-world evaluation cannot support the claim that slow changes are detected. The unified ground-truth annotations do not distinguish slow from fast changes, the reported F1=0.8126 is obtained after per-scene threshold selection with no protocol or held-out tuning, and the statement that detections from either branch are considered valid means the metric cannot attribute performance to the optical flow branch. Please provide separate ground-truth maps for slow and fast changes (or at least disjoint subsets), fix the threshold-selection protocol, and report baseline comparisons (e.g., the CD branch alone, the flow branch alone, and a frame-differencing method) on the same real-world data.","section":"4.3"},{"comment":"The model naming is inconsistent: the text states that Flow-CDNet utilizes SpyNetC2, while Table 1 lists Flow-CDNet-S (SpyNet+CDNet) separately and later sections identify the proposed Flow-CDNet as RAFT-based. This ambiguity makes it unclear which variant is the proposed method. In addition, Table 1 compares only single-task baselines and the authors' own variants; no existing change detection or optical flow network is compared on the same data, so the claim of outperforming existing methods is not demonstrated. Please clarify the naming and add comparisons to strong off-the-shelf baselines (e.g., DASNet, ChangeViT, GMFlow, FlowFormer).","section":"4.2, Table 1"},{"comment":"The Tversky term in Eq. (4) is the Tversky index, not the Tversky loss; minimizing this quantity would drive the true-positive term masked_gt toward zero, degenerating the change-detection output. The standard Tversky loss is 1 - (masked_gt / (masked_gt + α·wrong_classified + β·unmasked_gt)). Please correct the equation and explicitly state the optimization direction (minimization or maximization) used in training.","section":"3.2, Eq. (4)"},{"comment":"The definitions of EPE and FEPE are imprecise: Eq. (9) defines EPE as the square root of the absolute difference of flow vectors rather than the Euclidean norm of their difference, and Eq. (11) leaves the perturbation ε unspecified. The numerical values in Table 1 (F1=0.892, mEPE=1.027, FEPE=0.869) are approximately consistent with ε=0, but this is never stated. Since FEPE is used as the ranking metric, please define it rigorously, report the value of ε and its sensitivity, and use the standard endpoint-error definition.","section":"3.3, Eqs. (9)-(11)"},{"comment":"The Flow-Change synthetic dataset models slow changes as rigid displacements of PASCAL VOC objects on FlyingChairs backgrounds, which is not representative of the gradual, non-rigid deformation (e.g., soil movement or crack growth) described in the introduction. The real-world dam images lack flow ground truth, leaving the synthetic-to-real transfer for deformation unquantified. Please either temper the claims about detecting gradual deformation or add non-rigid deformation to the synthetic data and provide quantitative transfer evidence on real images.","section":"4.1"}],"minor_comments":[{"comment":"In the ablation paragraph, 'Flow-CDNet achieves a higher mEPE metric' should read 'lower mEPE', since a smaller mEPE indicates better optical flow estimation.","section":"4.2"},{"comment":"The loss weighting coefficient is denoted ψ in Eq. (5) but φ in Section 4.2; please unify the notation.","section":"3.2, 4.2"},{"comment":"Eqs. (9) and (10) should use explicit vector norms; the current notation is ambiguous and does not match standard definitions of endpoint error.","section":"3.3"},{"comment":"Please clarify how the fast-change objects are placed (e.g., appearing only in the second image) and how the binary change labels are computed, and provide a public link to the dataset and code to facilitate reproducibility.","section":"4.1"},{"comment":"The 'adaptive mask mechanism' is described only qualitatively; please specify its inputs, form, and role in the network, or omit the term if it is not a separate component.","section":"3.1.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's evaluation is almost entirely internal: self-built dataset, self-constructed variants, and a small real-world set with unified labels. The naming inconsistency between SpyNetC2 and RAFT as the main backbone raises concerns about which configuration was actually tested. These issues are fixable in a major revision, but the current evidence does not support the strong claims of simultaneous detection and real-world generalization."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes a genuinely new task: one network that simultaneously detects slow (displacement) and fast (appearance/disappearance) changes in bitemporal imagery, using a RAFT-style optical flow branch and a ResNet/PSP change detection branch. The task formulation is the real novelty; the architecture is a reasonable engineering combination, and the synthetic Flow-Change dataset plus the FEPE metric are new, if in-house. The ablation in Table 2 is the most convincing part: adding the CD branch to the flow branch drops mEPE from 2.409 to 1.027, and adding flow to the CD branch raises F1 from 0.753 to 0.892 on their synthetic test set. That supports mutual enhancement within that distribution, and the result is not present in the cited literature.\n\nThe soft spots are concentrated in evaluation. Every headline comparison runs only against self-constructed variants on a self-built dataset; there is no comparison with actual state-of-the-art CD or optical flow methods on standard benchmarks. The real-world validation in Section 4.3 is the weakest link: 20 image pairs, no flow ground truth, unified change labels that do not distinguish slow from fast, and an F1 of 0.8126 obtained after per-scene threshold selection. I agree with the stress-test concern: that number cannot substantiate the slow-change claim, because a change-detection-only model could match or beat it if the labels are dominated by appearance changes. The flow branch's real-world contribution is unverified. Also, the text has internal inconsistencies (SpyNetC2 vs RAFT as backbone, psi vs phi for the loss weight, Eq 9 missing the squared norm in EPE), and no code or data are released.\n\nThese are addressable weaknesses, not fatal flaws. The central engineering idea is sound and the synthetic ablation is internally coherent. The paper deserves a serious referee, but a referee should require stronger evaluation: comparisons on standard benchmarks, a real-world protocol that separates slow and fast detections or at least reports qualitative evidence tied to known deformation sites, and ideally some flow ground truth for real scenes. I would not desk-reject; I would send it out with a request for major revisions.","headline":"A plausible dual-branch joint slow/fast change detection idea with a coherent synthetic ablation, but real-world evidence is too weak to support the flagship claim of detecting gradual deformation.","tokens_in":16745,"tokens_out":2050,"would_cite":false,"duration_ms":24199,"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":"A single dual-branch network, Flow-CDNet, claims to detect both slow displacements and sudden changes in bitemporal images, outperforming flow-only and change-only baselines on a synthetic benchmark and generalizing to real dam-bank…","keywords":["change detection","optical flow","bitemporal images","slow change","fast change","dual-branch network","FEPE metric","synthetic dataset"],"falsifier":"Take a set of real slope or dam image pairs with independently measured dense displacement fields, for example from survey markers or controlled laboratory deformation, run the trained Flow-CDNet without fine-tuning, and compare its mean end-point error in slowly deforming regions against a flow-only RAFT baseline; if the dual-branch model does not beat or match that baseline there, or if its change masks miss gradually deforming areas, the claimed simultaneous slow and fast change detection is not supported.","tokens_in":1801,"feed_emoji":"🛰️","tokens_out":2123,"duration_ms":92739,"temperature":0.7,"pith_summary":"Change detection in monitoring imagery usually targets fast changes, such as objects appearing or disappearing between two pictures. This paper takes on the harder, practically relevant problem of also catching slow changes, the small displacements of soil, rock, or structure that often precede failures like slope slides or dam damage. The paper's central claim is that one dual-branch network, Flow-CDNet, can detect both kinds at once: one branch estimates dense optical flow, and the other uses that flow to produce a binary change mask. On the paper's synthetic Flow-Change dataset, the joint model reaches an F1 score of 0.892, a mean end-point error of 1.027, and an FEPE score of 0.869, beating flow-only, change-only, and combined baselines, and it transfers to 20 real dam-bank image pairs with an F1 score of 0.8126.","feed_headline":"One network detects slow shifts and sudden changes in images","feed_subtitle":"Dual-branch model pairs optical flow with change masks, beats flow-only and change-only baselines, and transfers to dam imagery.","key_machinery":"The machine that carries the argument is the dual-branch coupling between optical flow and change detection. The optical-flow branch uses a RAFT-style architecture with a four-level correlation pyramid and convolutional GRU updates to estimate a dense displacement field; the change-detection branch warps the second image by that flow, takes the absolute difference with the first image, applies an adaptive mask, and pushes the result through a ResNet50 backbone with pyramid pooling to produce a binary change map. This coupling is supervised by a composite loss, L2 flow error on pixels outside fast-change regions plus Tversky loss on the change mask, and is measured by FEPE, a single score defined as F1 divided by the sum of mean end-point error and a small epsilon, so a model must be good at both tasks to rank well.","core_discovery":"The paper introduces Flow-CDNet, a unified framework in which a pyramid optical-flow branch and a ResNet-based binary change-detection branch are trained together so that slow changes, where an object is present in both images but shifted or deformed, and fast changes, where an object appears in only one image, are detected simultaneously. The flow branch is built on RAFT-style feature extraction, a multi-scale 4D correlation volume, and iterative update, and it feeds the change branch with a motion-compensated warped image, an absolute difference map, and an adaptive mask; the change branch then outputs a binary map under a combined loss of L2 flow error, masked to exclude fast-change regions, and Tversky segmentation loss. The authors report that Flow-CDNet reaches F1 0.892 and a mean end-point error of 1.027 on the synthetic Flow-Change benchmark, giving FEPE 0.869, and their ablation shows the two branches improve each other: adding the flow branch raises F1 from 0.753 to 0.892, while adding the change branch lowers the flow error from 2.409 to 1.027. They further claim that the pretrained model, without fine-tuning, detects both abrupt collapses and gradual surface deformation in real dam-bank imagery, reaching F1 0.8126 after per-scene thresholds. On its own terms, the contribution is a working demonstration that a single network can produce both a usable dense flow field and a usable change mask for monitoring scenarios.","pith_inferences":["The decisive next test is not a new network but a new dataset: real slope or dam image pairs with independently measured subpixel displacement fields, because the synthetic slow changes are rigid object shifts and do not exercise genuine gradual deformation.","A useful ablation the paper does not report is freezing a pretrained flow branch and training only the change head; that would separate the benefit of motion-aligned inputs from the benefit of joint gradient updates, pinning down where the mutual improvement comes from.","The FEPE ratio should be treated cautiously as a composite: a model with a slightly worse F1 but a much smaller flow error can outrank a more balanced model, so applications should inspect the two components as well as the combined score.","An extension this framework invites is predicting a continuous change-magnitude map instead of a binary mask, which would let slow and fast changes live on the same output and make the method directly usable for early-warning thresholds."],"forward_implications":["A monitoring system could run one inference per image pair and receive both a dense displacement field and a binary change mask, so gradual precursors and sudden events are flagged together.","Because the two branches improve each other, with F1 rising from 0.753 to 0.892 and flow error falling from 2.409 to 1.027, architectures built for either task alone may gain from adding the other task's head and loss.","The FEPE metric gives a single ranking for models that output both flow and change maps, making slow-plus-fast performance comparable rather than reporting two separate numbers.","The reported transfer to real dam-bank pairs without fine-tuning suggests that a model trained on synthetic flow-plus-change data can generalize to unseen monitoring sites, at least for the change types represented in the test set."],"supporting_citations":[{"why":"Supplies the RAFT optical-flow backbone used in the flow branch, including the recurrent all-pairs field transforms, correlation pyramid, and iterative updates.","marker":"[5]"},{"why":"Provides the FlyingChairs background image pairs and optical-flow ground truth used to construct the synthetic Flow-Change dataset.","marker":"[14]"},{"why":"Supplies the PASCAL VOC object regions that are randomly transformed and pasted onto backgrounds to simulate fast changes and to define change labels.","marker":"[60]"},{"why":"PSPNet is used as the change-detection baseline and design reference, providing the spatial pyramid pooling structure adopted by the CD branch.","marker":"[8]"},{"why":"Introduces the Tversky loss that supervises the binary change mask in the combined loss function.","marker":"[59]"},{"why":"SpyNet serves as one of the compared optical-flow backbones used to construct a Flow-CDNet variant in the experiments.","marker":"[1]"},{"why":"LiteFlowNet serves as another compared optical-flow backbone used to construct a Flow-CDNet variant in the experiments.","marker":"[4]"}],"fun_headline_variants":["Dual-branch network catches slow and fast changes in images","Flow-CDNet: one network for gradual shifts and sudden changes","Pairing optical flow with change masks boosts bitemporal detection","Detect both deformation and abrupt events with a single model","From slow slides to fast breaks: unified change detection"],"cache_read_input_tokens":18688,"weakest_assumption_plain":"The paper's central comparison rests on the assumption that its synthetic dataset, made by pasting transformed objects from one image collection onto optical-flow pairs from another, faithfully represents real slow and fast changes; the slow changes it contains are simple object shifts, not gradual small-scale ground movements.","fun_headline_variants_meta":{"raw":{"variants":["Dual-branch network catches slow and fast changes in images","Flow-CDNet: one network for gradual shifts and sudden changes","Pairing optical flow with change masks boosts bitemporal detection","Detect both deformation and abrupt events with a single model","From slow slides to fast breaks: unified change detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000235,"raw_usage":{"total_tokens":1562,"prompt_tokens":1072,"completion_tokens":490,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":688,"completion_tokens_details":{"reasoning_tokens":408}},"tokens_in":688,"tokens_out":490,"duration_ms":5785,"temperature":1.0,"reasoning_tokens":408,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:33:54.776768+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of real slope or dam image pairs with independently measured dense displacement fields, for example from survey markers or controlled laboratory deformation, run the trained Flow-CDNet without fine-tuning, and compare its mean end-point error in slowly deforming regions against a flow-only RAFT baseline; if the dual-branch model does not beat or match that baseline there, or if its change masks miss gradually deforming areas, the claimed simultaneous slow and fast change detection is not supported.","supporting_citations":[{"cited_title":"Raft: Recurrent all-pairs field transforms for optical flow","cited_arxiv_id":null,"evidence_quote":"Supplies the RAFT optical-flow backbone used in the flow branch, including the recurrent all-pairs field transforms, correlation pyramid, and iterative updates."},{"cited_title":"Flownet: Learning optical flow with convolutional networks","cited_arxiv_id":null,"evidence_quote":"Provides the FlyingChairs background image pairs and optical-flow ground truth used to construct the synthetic Flow-Change dataset."},{"cited_title":"The pascal visual object classes challenge: A retrospective","cited_arxiv_id":null,"evidence_quote":"Supplies the PASCAL VOC object regions that are randomly transformed and pasted onto backgrounds to simulate fast changes and to define change labels."},{"cited_title":"Tversky loss function for image segmentation using 3d fully convolutional deep networks","cited_arxiv_id":null,"evidence_quote":"Introduces the Tversky loss that supervises the binary change mask in the combined loss function."},{"cited_title":"Optical flow estimation using a spatial pyramid network","cited_arxiv_id":null,"evidence_quote":"SpyNet serves as one of the compared optical-flow backbones used to construct a Flow-CDNet variant in the experiments."},{"cited_title":"Liteflownet: A lightweight convolutional neural network for optical flow estimation","cited_arxiv_id":null,"evidence_quote":"LiteFlowNet serves as another compared optical-flow backbone used to construct a Flow-CDNet variant in the experiments."}],"review_version":1}