{"id":"5e930883-7ff9-475b-9263-cb890c7f47af","arxiv_id":"2507.18653","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A modular branching method with contrastive routing preserves source lane detection performance while adapting to target distributions with fewer trained parameters.","lead":"This paper studies lane detection models that fail when moving between datasets, and proposes keeping a frozen source model while adding small fine-tuned branches for each new domain, routed by a contrastive classifier. A generalist might care because it is a practical recipe for adapting perception models to new environments without forgetting old ones, using fewer parameters than training separate systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3's Avg F1 is an unweighted mean of per-distribution branch scores, not a measured end-to-end routed F1, and the routing point is selected from the test table—so the parameter-efficiency claim lacks a valid joint evaluation.","rationale":"The reader's weakest assumption focuses on the SCL classifier's near-perfect accuracy and the missing end-to-end routed F1. I agree that this is load-bearing: the paper's central claim is about the deployed routing system, but Table 3 gives only per-distribution branch F1s averaged together, and the joint system in Eq. (9) is never evaluated. My additional concern is the test-set selection of the routing point: the '41% of parameters' result uses Routing@N+H for CULane and Routing@B(k=2)+N+H for AssistTaxi, both highlighted after inspecting test performance, so the reported efficiency may not be a pre-specified design choice. Both concerns are addressable with a validation-split rule and a joint routed evaluation, so the evidence supports retaining the reader's CONDITIONAL verdict rather than moving to ACCEPT or REJECT. If the proposed end-to-end evaluation confirms the arithmetic-mean F1 and the validation-selected configuration matches Table 3, the central claim would be substantially strengthened; otherwise, the near-optimal and parameter-efficiency statements would need to be weakened or re-scoped.","tokens_in":19071,"tokens_out":5930,"duration_ms":65284,"concrete_test":"Run the full inference pipeline end-to-end: train the SCL classifier on training splits, freeze all branches, pass a mixed test set (CULane/CurveLanes/AssistTaxi) through the classifier and the corresponding branch, and compute one F1 over the routed outputs, with per-class confusion counts. Separately, choose Routing@X (H, N+H, or B(k=2)+N+H) using only a validation split, then report the validation-selected configuration's test F1 and parameter count against the test-selected numbers in Table 3. If the end-to-end routed F1 differs from the arithmetic mean by more than about 1 F1 point, or if validation selection picks a different routing point than the highlighted one, the near-optimal and parameter-efficiency claims weaken.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the Abstract is that the inference-time framework achieves near-optimal F1 with fewer parameters. To support this, Eq. (9) defines the routed output, but the paper never evaluates the joint system y = f_{theta'_hat_d(x)}(x). Section 4.4 reports only the SCL classifier's accuracy (99.6/99.7/99.9%) and the 0.31 ms routing overhead. Table 3 reports separate columns F1 Src, F1Curve, F1Assist and their arithmetic average; these are branch-specific test scores, not the result of routing a mixed test set through the classifier plus branch, since the source F1 is constant across routing configurations. A misclassified input is sent to a branch adapted to another distribution, and Tables 1/5 show cross-branch performance can be near zero (e.g., a CULane ERFNet N+H model fine-tuned on AssistTaxi has source F1 0.8 while target F1 is 86.6). Even at 99.6% classifier accuracy, the joint F1 is not demonstrated. In addition, the comparison underlying '41% of parameters' chooses Routing@N+H for CULane and Routing@B(k=2)+N+H for AssistTaxi, both highlighted only after examining test-set Table 3; no validation-based rule is specified. Thus the headline parameter-efficiency and near-optimal F1 claims rest on an unevaluated joint inference path and a test-selected routing configuration.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a modular branching framework for adapting CLRerNet lane detection models to multiple distributions while avoiding catastrophic forgetting. The authors first train on a source dataset, then create cloned branches for each target and fine-tune only selected components (bias, head, neck+head, partial backbone, full) based on a component-wise analysis across CULane, CurveLanes, and AssistTaxi with DLA-34, ResNet-18, and ERFNet backbones. At inference, a supervised contrastive learning (SCL) classifier estimates the input's distribution and routes it to the corresponding branch. Table 3 reports per-configuration F1 averages and parameter counts, and the paper claims near-optimal F1 with fewer parameters than training separate models.","tokens_in":19423,"tokens_out":5892,"duration_ms":59832,"significance":"The component-wise forgetting analysis is a potentially useful empirical contribution: it covers three backbones and three datasets, reports source and target F1 drops/gains, and gives transparent parameter counts. The SCL routing overhead (0.31 ms) and near-perfect per-dataset classification accuracy are also promising. However, the central claim of the paper—that the inference-time routing framework achieves near-optimal F1 with fewer parameters—is not yet demonstrated, because the joint system is never evaluated and the routing-depth choice is made from the test table. The strengths are real but do not currently support the abstract's headline claim.","major_comments":[{"comment":"The central evaluation is missing: Eq. (9) defines the routed output, but the paper never computes an end-to-end F1 on a mixed test set. Table 3's F1Avg is the unweighted arithmetic mean of the three per-distribution branch F1s; because F1 Src is constant by construction and the SCL classifier is evaluated separately in §4.4, F1Avg does not reflect misrouting. With reported SCL accuracy of at most 99.9%, on a mixed test set roughly 0.1–0.4% of inputs would be sent to a branch whose target F1 can be near zero (e.g., an AssistTaxi N+H branch on a CULane image has F1 0.8 in Table 5). The abstract's 'near-optimal F1-scores' is therefore not directly supported until the joint system is evaluated (e.g., a mixed test stream with the confusion matrix and the routed F1).","section":"§4.4, Table 3, Eq. (9)"},{"comment":"The highlighted routing configurations are selected based on the test-set F1Avg. For CULane source, Routing@N+H (F1Avg 80.8) is claimed best even though Routing@B(k=2)+N+H and Routing@B+N+H both reach 83.6; for AssistTaxi source, Routing@B(k=2)+N+H (79.2) is chosen over Routing@B+N+H (82.3) and Routing@N+H (43.6). No validation-based rule for choosing the routing depth is given, so the '41% of parameters' headline compares a configuration that was picked after inspecting the test table against a full fine-tuning baseline. Please either select via a validation split or report all routing options without asserting a single best configuration.","section":"Table 3"},{"comment":"The supervised contrastive classifier is central to routing but its training is underspecified: Eq. (5) defines the loss, yet the paper does not state the encoder/MLP architecture, optimizer, learning rate, epochs, batch size, temperature, augmentation, or how the centroids in Eq. (6) are computed. These details are needed to reproduce the claimed 99.6–99.9% accuracy and to judge whether the classifier would transfer to a genuinely mixed test stream rather than being fitted to the same three known splits.","section":"§3.4"},{"comment":"Table 2 reports AssistTaxi source F1 = 0.0 for every fine-tuning configuration, including B+N+H, while Table 3 reports F1 Src = 94.3 for the corresponding Routing@B+N+H. If Table 2 describes standard fine-tuning without branch preservation, the text should say so explicitly; if it describes the proposed branching, it contradicts the 'no source forgetting' claim. This apparent inconsistency needs to be resolved.","section":"Table 2 vs. Table 3"}],"minor_comments":[{"comment":"Equation (7) contains a typographical error in the subscript: the argmin is written over 'k in 1,...,K ; |...|' with a semicolon inside the expression; it should be 'argmin_k ||g_psi(enc_phi(x)) - mu_k||_2^2'.","section":"Eq. (7)"},{"comment":"The caption says F1 Src is constant across configurations and is 'denoted with  ' ', but the table entries for F1 Src appear blank; please use a visible placeholder or repeat the values to avoid confusion.","section":"Table 3 caption"},{"comment":"The 'Relative' column treats Routing@B+N+H as 100%, but the surrounding text says the comparison is against 'fine-tuning fully separate models (treated as baseline)'; clarify whether the baseline is separate full models or the branched full-routing configuration.","section":"Table 3, Relative column"},{"comment":"No error bars or multiple-seed results are reported; given that some Table 3 differences are only 2–3 F1 points (e.g., 80.8 vs. 83.6), it would be helpful to state whether these gaps are stable across seeds.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The central evaluation gap is significant: the headline claim rests on an unevaluated joint inference path and a test-selected routing configuration. If the authors provide a proper end-to-end routed F1 evaluation on a mixed test set and a validation-based selection rule, the paper could become acceptable. The AssistTaxi source inconsistency between Table 2 and Table 3 should also be resolved before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plain take: the component-wise fine-tuning study is the real substance, and it is mostly sound. The contrastive-routing wrapper is a plausible idea, but the headline claim — near-optimal F1 at 41% of parameters — is not demonstrated, because Table 3 is an arithmetic mean of per-branch F1 scores, not a measured F1 of the routed system.\n\nWhat is new: a systematic comparison of bias-only, head-only, neck+head, partial-backbone, and full fine-tuning for CLRerNet on CULane, CurveLanes, and AssistTaxi, across DLA-34, ResNet-18, and ERFNet. The finding that moderate shift (CULane to CurveLanes) is handled well by N+H while severe shift (AssistTaxi to CULane) needs partial backbone tuning is useful and is supported by the tables. Freezing the source branch is a clean way to avoid forgetting, and the SCL classifier accuracy (99.6-99.9%) with 0.31 ms overhead is honestly reported. The limitations section acknowledges the branching overhead.\n\nThe soft spots, in proportion:\n1. No end-to-end routed evaluation. Equation (9) defines the joint output, but the paper never computes F1 on a mixed stream using predicted route labels. Table 3 keeps F1 Src constant across configurations, so it cannot be a routed result. Even at 99.6% classifier accuracy, roughly 0.4% of inputs go to the wrong branch, and cross-branch F1 can be near zero (Tables 1 and 5 show source F1 dropping to 0.8 for an ERFNet N+H branch after AssistTaxi fine-tuning). The actual loss depends on which samples are misrouted, and we don't know it.\n2. The routing point is selected post hoc. N+H is highlighted for CULane and B(k=2)+N+H for AssistTaxi after examining which row has the best Avg F1 in Table 3. No validation-based rule is given. So the \"near-optimal\" claim is partly fitted to this test set.\n3. No error bars or repeated-run variance. Combined with the 3-epoch fine-tuning budget, the margin between Routing@N+H and Routing@B(k=2)+N+H (about 2.8 F1 for CULane source) may not be stable.\nMinor: the SCL training details (optimizer, epochs, temperature) are missing, and Table 2 lists source F1 as 0.0 after every fine-tuning configuration, which looks like an artifact; the prose elsewhere doesn't match that.\n\nWho this is for: people working on transfer learning for lane detection, and anyone interested in how evaluation gaps can inflate a domain-adaptation claim. The component-wise data is a reasonable reference even if the routing framework needs rework.\n\nRecommendation: send to peer review, but require a true joint evaluation on a mixed test set, a validation-based rule for choosing the routing point, and error bars. The component analysis alone is worth referee time.","headline":"Solid component-wise fine-tuning study; the routing claim outruns the evidence — Table 3 is a branch-score average, not an end-to-end routed F1.","tokens_in":19930,"tokens_out":4410,"would_cite":true,"duration_ms":44380,"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":"The paper claims that modular per-distribution branches plus contrastive routing adapt lane detectors to new datasets while preserving source performance and using fewer parameters than full fine-tuning.","keywords":["lane detection","distribution shift","catastrophic forgetting","fine-tuning","supervised contrastive learning","model routing","parameter-efficient adaptation"],"falsifier":"Run the full routing pipeline end-to-end on the three test sets and compare the actual routed F1 with the oracle-routed F1 reported in Table 3; if the 99.6–99.9% classifier accuracy does not translate into matching end-to-end scores, the claim of near-optimal routed performance fails. A sharper test is to feed in a held-out fourth distribution that is visually close to one of the three, for example night-time CULane, and check whether the classifier forces it to the nearest centroid and the wrong branch drops F1.","tokens_in":18921,"feed_emoji":"🛣️","tokens_out":5958,"duration_ms":58688,"temperature":0.7,"pith_summary":"Lane detection models trained on one dataset lose their original skill when fine-tuned on another, even when both are road scenes; the paper calls this catastrophic forgetting and shows it can be severe enough to drop source F1 from 80 to near zero. The proposed remedy is to keep the source model intact and add cloned branches for each new distribution, fine-tuning only the components that matter—the detection head and feature neck for mild shifts, and deeper backbone layers as well for large shifts. At inference, a supervised contrastive classifier decides which distribution an image comes from and routes it to the corresponding branch. The paper claims this reaches near-optimal F1 on all three datasets it tests while using as little as 41% of the parameters of training separate full models, with source performance unchanged.","feed_headline":"Route lane images by distribution to stop catastrophic forgetting","feed_subtitle":"Cloned neck+head branches preserve source accuracy and cut parameters to 41% of full fine-tuning.","key_machinery":"The machinery is a modular version of the anchor-based detector CLRerNet, split into backbone B, neck N, and head H, where N and H are cloned and fine-tuned per target while B stays frozen in the light setting, or the last k backbone layers are included in the clone for severe shifts. Inference is driven by a supervised contrastive learning (SCL) classifier: a frozen ResNet encoder with a projection MLP is trained with the SupCon loss so that embeddings of the same distribution cluster tightly, then each distribution is represented by a centroid and an input is assigned to the nearest centroid, selecting the matching branch. The routing point X in 'Routing@X' is the layer at which per-distribution specialization starts; layers before X are shared, layers from X onward are cloned.","core_discovery":"The core claim is that adaptation to a new lane-detection distribution and retention of the source distribution need not trade off against each other if the network is reorganized into frozen shared modules plus per-distribution branches. Across three datasets—CULane, CurveLanes, and AssistTaxi—the paper shows that fine-tuning only the neck and head preserves most target accuracy for moderate shifts, while for a severe shift such as airport taxiways from a car-lane base, unfreezing the last two backbone layers plus neck and head is required. With CULane as source, routing at neck+head gives an average F1 of 80.8 with 41% of full fine-tuning's parameters, against 83.6 for full adaptation. With AssistTaxi as source, routing at partial backbone gives 79.2 with 96% of parameters, against 82.3 for full adaptation. Because the source branch is never updated, source F1 is retained by construction. A supervised contrastive learning classifier trained on the three distributions reports 99.6–99.9% accuracy and routes each input to the matching branch.","pith_inferences":["A testable extension is to predict the needed routing depth from an embedding-space distance between source and target, such as the SCL centroid distance, rather than from empirical sweeps; the paper does not make this link.","The closed-set assumption is the natural boundary: an unknown distribution will be forced to the nearest known centroid, so adding a rejection rule or an 'unknown' class to the SCL classifier would be needed for open-world deployment.","The same modular branch-and-route scheme likely transfers to other anchor- or row-wise lane detectors whose heads encode positional priors, since only the component roles are used, but this remains to be demonstrated.","As more target distributions are added, branch count and memory grow linearly; distilling routed predictions back into a single shared network, which the paper only suggests as future work, would trade a small accuracy loss for constant parameter growth."],"forward_implications":["Source performance is preserved exactly, because the source branch parameters are never modified during adaptation.","For moderate shifts such as CULane to CurveLanes or AssistTaxi, neck+head branches match most of the full fine-tuning gain with only about 41% of the parameters (6.8M versus 16.7M on ERFNet).","For severe shifts such as AssistTaxi to road lanes, shallow branches fail, with average F1 of 32.5–43.6, while partial-backbone branches reach 79.2 average F1 at 96% of the parameters.","The 0.31 ms added by contrastive routing on top of 9.82 ms base inference keeps the system real-time on an RTX 4080.","The component-wise observations give a heuristic for choosing the branch point: the larger the distribution change, the deeper the fine-tuned modules must go."],"supporting_citations":[{"why":"CLRerNet is the base lane detector whose backbone, neck, and head are modularized into shared and per-distribution branches.","marker":"[20]"},{"why":"Supervised contrastive loss is the training objective for the distribution classifier used to route inputs.","marker":"[28]"},{"why":"AssistTaxi provides the airport-taxiway distribution that causes near-zero zero-shot F1 and defines the hardest shift.","marker":"[10]"},{"why":"CULane is the large-scale urban lane dataset used as a source or target distribution in the routing experiments.","marker":"[46]"},{"why":"CurveLanes is the curved-lane dataset used as an additional source or target distribution.","marker":"[58]"},{"why":"ERFNet is the lightweight backbone used for the routing experiments and the real-time timing measurements.","marker":"[51]"},{"why":"Feature pyramid networks define the neck module whose multi-scale fusion is fine-tuned for target adaptation.","marker":"[34]"}],"fun_headline_variants":["Adaptive lane routing preserves source accuracy","Fine-tune only the neck and head to avoid forgetting","Contrastive routing cuts parameters to 41%","Branch-based adaptation beats full fine-tuning","Route by distribution to keep lane detection sharp"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system assumes every test image belongs to exactly one of the known training distributions and that the contrastive classifier always sends it to the right branch; a single wrong route into the wrong branch can produce near-zero F1 for extreme shifts.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive lane routing preserves source accuracy","Fine-tune only the neck and head to avoid forgetting","Contrastive routing cuts parameters to 41%","Branch-based adaptation beats full fine-tuning","Route by distribution to keep lane detection sharp"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000108,"raw_usage":{"total_tokens":1027,"prompt_tokens":905,"completion_tokens":122,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":53}},"tokens_in":521,"tokens_out":122,"duration_ms":2147,"temperature":1.0,"reasoning_tokens":53,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:59:47.110823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full routing pipeline end-to-end on the three test sets and compare the actual routed F1 with the oracle-routed F1 reported in Table 3; if the 99.6–99.9% classifier accuracy does not translate into matching end-to-end scores, the claim of near-optimal routed performance fails. A sharper test is to feed in a held-out fourth distribution that is visually close to one of the three, for example night-time CULane, and check whether the classifier forces it to the nearest centroid and the wrong branch drops F1.","supporting_citations":[{"cited_title":"Clrernet: improving con- fidence of lane detection with laneiou","cited_arxiv_id":null,"evidence_quote":"CLRerNet is the base lane detector whose backbone, neck, and head are modularized into shared and per-distribution branches."},{"cited_title":"Assisttaxi: A comprehensive dataset for taxiway analysis and autonomous operations","cited_arxiv_id":null,"evidence_quote":"AssistTaxi provides the airport-taxiway distribution that causes near-zero zero-shot F1 and defines the hardest shift."},{"cited_title":"Spatial as deep: Spatial cnn for traffic scene understanding","cited_arxiv_id":null,"evidence_quote":"CULane is the large-scale urban lane dataset used as a source or target distribution in the routing experiments."},{"cited_title":"Curvelane-nas: Unifying lane- sensitive architecture search and adaptive point blending","cited_arxiv_id":null,"evidence_quote":"CurveLanes is the curved-lane dataset used as an additional source or target distribution."},{"cited_title":"Erfnet: Efficient residual factorized convnet for real-time semantic segmentation","cited_arxiv_id":null,"evidence_quote":"ERFNet is the lightweight backbone used for the routing experiments and the real-time timing measurements."}],"review_version":1}