{"id":"5b52d01d-f4ea-4542-9926-1031fd0e43f5","arxiv_id":"2511.06720","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A relative-energy score over paired known/unknown logits plus road-point perturbation (Point Raise) achieves state-of-the-art LiDAR OOD detection on STU and SemanticKITTI.","lead":"This paper gives self-driving LiDAR perception a way to flag objects never seen in training, by scoring each point on the relative evidence for 'unknown' versus 'known' classes and training on cheaply synthesized fake obstacles. It reports large gains over prior methods on the STU and SemanticKITTI benchmarks, with a roughly four-fold reduction in false positives at high recall.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"STU test table omits UEM, the validation SOTA; without it the 'state-of-the-art' claim is not fully supported.","rationale":"The reader's weakest_assumption is the Point Raise coverage of real OOD geometry. While that is a legitimate generalization concern, the empirical results on held-out test sets already provide partial support for the transfer of the synthetic outliers to the benchmark OOD. A more direct threat to the central claim is that the STU test comparison omits UEM, the strongest baseline from validation. This is a concrete evidence gap: the paper's 'state-of-the-art' claim on the primary benchmark rests on a table that excludes the most competitive method. The reader did list this as a reproducibility condition, so my concern is not entirely new, but I elevate it to the load-bearing issue. The fix is straightforward and uses the authors' already-existing UEM reproduction. If UEM matches its validation numbers on test, REL still appears better, but the magnitude of the improvement is unclear; if UEM is closer on test, the claim weakens. Thus the verdict remains CONDITIONAL pending this specific comparison.","tokens_in":14028,"tokens_out":9700,"duration_ms":97935,"concrete_test":"Reproduce UEM (Nayal et al. 2025) on the STU test split using the same backbone and training setup as the authors' validation reproduction, and add its AUROC/FPR@95/AP to Table 2. The decisive comparison: if UEM's test AUROC ≥ 96.26 or FPR@95 ≤ 21.69, the claimed 'state-of-the-art' on the primary benchmark is unsupported; if UEM is clearly worse (e.g., FPR@95 > 30), the claim is corroborated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 2 (STU test set) reports point-level AUROC/FPR@95/AP for Deep Ensemble, MC Dropout, Max Logit, Void Classifier, RbA, and Ours, but omits UEM [37], which is the strongest training-based baseline on the validation set (Table 1: UEM* AUROC 95.80, FPR@95 26.37 vs Ours 97.85/9.60). The paper states test results are 'taken directly from [39]', but UEM is not from [39]; it was reproduced by the authors for validation only. Since the abstract and Fig. 1 make a global 'state-of-the-art, highest AUROC, lowest FPR@95' claim, leaving out the strongest competitor on the held-out test split is a direct evidence gap. If UEM's test performance is close to its validation numbers (≈95.8 AUROC / 26.4 FPR), REL's test 96.26/21.69 is still superior, but the margin is much narrower than the 'large margin' claimed; if UEM is within noise, the SOTA claim is not established. This concern is independent of the Point Raise proxy assumption: it concerns the completeness of the evidence presented for the central claim, and it is easily resolved by running UEM on the STU test set with the same protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Relative Energy Learning (REL), a point-wise OOD detection framework for LiDAR semantic/panoptic segmentation. REL defines a relative energy margin ΔE = log(p_neg/p_pos) between auxiliary negative logits and in-distribution positive logits, trains a lightweight OOD projector with a weighted logistic loss, and introduces Point Raise, a geometry-aware synthesis that contracts and lifts patches of road points to create pseudo-OOD training samples. The method is evaluated on the STU benchmark and SemanticKITTI, reporting state-of-the-art AUROC and FPR@95 over prior post-hoc and training-based baselines. The derivation in Eqs. (4)-(6) is correct, and the held-out STU test set is used for the main test-table results. However, the evidence for the headline claim is weakened by the omission of the strongest validation baseline from the test table, the absence of variance estimates, and under-reported per-class closed-set degradation.","tokens_in":14290,"tokens_out":5146,"duration_ms":47657,"significance":"If the reported results hold, REL is a simple and practical contribution: the relative-energy score is a natural log-odds formulation, avoids margin and temperature hyperparameters, and Point Raise is a lightweight alternative to external synthetic OOD sources. The reported gains are large (STU test AUROC 96.26 vs Deep Ensemble 86.74 and FPR@95 21.69 vs 58.05; SemanticKITTI AUROC 96.76 vs UEM 93.15), and the ablations isolate the contribution of the REL objective and the Point Raise augmentation. Strengths of the paper include the transparent derivation of ΔE, the use of a held-out STU test split, and a closed-set quality comparison. The main weaknesses are evidentiary: the STU test table omits UEM, the strongest reproduced baseline; no multi-seed or error-bar analysis is provided despite stochastic augmentation; and the per-class closed-set tables show severe drops on rare classes that are not discussed in the main text. These issues do not invalidate the method, but they need to be addressed before the state-of-the-art claim can be considered fully supported.","major_comments":[{"comment":"The STU test-set table and Figure 1 omit UEM [37], which is the strongest training-based baseline on the validation set (Table 1: UEM* AUROC 95.80, FPR@95 26.37). The caption states test results are taken from [39], but UEM is not from [39]; it was reproduced by the authors only for validation. Since the abstract and Fig. 1 make a global state-of-the-art claim, leaving out the strongest competitor on the held-out test split is a direct evidence gap. Please run UEM on the STU test set with the same protocol and report it. If UEM's test performance is close to its validation numbers, the margin over the strongest baseline is much narrower than the paper's 'large margin' wording suggests.","section":"Table 2 / Fig. 1"},{"comment":"No error bars or multi-seed variance are reported in any table. Point Raise involves random sampling of road points, radii, heights, and the training procedure is stochastic, so single-run differences such as FPR@95 9.60 vs 21.69 may not be significant. Please provide mean and standard deviation over at least three seeds for the main results and ablations, or otherwise justify that the reported differences are stable.","section":"Tables 1-3 and 6-7"},{"comment":"The claim in Section 5.2 that REL preserves closed-set segmentation with 'minimal impact' is contradicted by the per-class results: bicycle PQ drops from 47.65 to 0.0 on STU, and motorcyclist PQ drops from 25.01 to 0.0 on SemanticKITTI. Overall PQ is comparable, but the collapse of these rare classes is a substantial degradation. Please either report and discuss this trade-off in the main text, or provide an explanation (e.g., class-frequency interaction with fine-tuning) and adjust the claim accordingly.","section":"Supplementary Tables 8 and 9"},{"comment":"Point Raise assumes that radially contracted and lifted road patches are a sufficient synthetic surrogate for real OOD objects. The paper offers cross-dataset AUROC as indirect evidence but no mechanistic argument that the synthetic geometry covers real anomalies (e.g., non-protruding, off-road, or overhead objects). The decision boundary learned from these synthetic protrusions may not transfer. A concrete test would be to stratify STU test anomalies by geometric properties (point count, protrusion, off-road placement) and report per-category performance, or to compare against an alternative synthetic OOD source such as LiON's ShapeNet-based outliers. This is a correctness-risk concern, not a circularity issue, and it directly affects the generalizability claim.","section":"Section 3.4 / Eq. (8)"}],"minor_comments":[{"comment":"The notation d_min/d_max is unclear; use d_min and d_max. Also clarify that the scaling s applies to the x,y coordinates only, and that the height perturbation h has per-point dimensionality matching the cluster size.","section":"Algorithm 1 / Eq. (8)"},{"comment":"The number of negative logits K is not explicitly defined in the method. It appears to equal the number of ID classes, but this should be stated clearly.","section":"Section 3.3"},{"comment":"Reference [16] has a typo: 'kuefeng Du' should be 'Kufeng Du'. There is also a typo in Section 2: 'effectively seperate' should be 'effectively separate'.","section":"References"},{"comment":"REAL, APF, and LiON use the Cylinder3D architecture while the other methods use Mask4Former. Since the comparison is across backbone architectures, this should be explicitly noted as a potential confound in the text.","section":"Table 3"},{"comment":"The 'None*' row means unlabeled points are used as auxiliary OOD data. This is important and should be explained in the main text, not only in the caption.","section":"Table 6"},{"comment":"The phrase 'our proposed baseline' in the supplementary qualitative section is ambiguous; use 'our method' for consistency.","section":"Supplementary, Sec. 7"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a clean derivation and a promising method. The main barrier to acceptance is the missing UEM test-set comparison, which is a straightforward experiment to run. The lack of variance reporting and the per-class closed-set collapses are also important. If these are addressed, the paper could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does one genuinely new thing: it trains an explicit OOD branch with a weighted logistic loss on the log-odds of grouped softmax, and it synthesizes pseudo-OOD points by radially contracting road patches and lifting them. The derivation (Eq. 4-6) is correct, and the empirical result on STU — validation AUROC 97.85 vs Deep Ensemble's 90.93, test AUROC 96.26 vs 86.74, plus strong cross-dataset numbers on SemanticKITTI — is impressive if it holds up. The ablations are honest: Relative Energy beats Hinge, VOS, and Dual Energy under the same synthesis, and the closed-set panoptic quality stays close to Mask4Former.\n\nThe soft spots are about evidence completeness, not the core idea. The biggest one is the missing UEM comparison on the STU test set. UEM is the strongest training-based baseline on validation (95.80 AUROC, 26.37 FPR@95), but Table 2, which reports test results taken from [39], doesn't include it. Since the abstract and Fig. 1 claim state-of-the-art with the lowest FPR, leaving out the most competitive baseline on the held-out split is a direct gap. If UEM's test numbers are close to its validation ones, REL still wins, but the margin is much thinner than advertised.\n\nSecond, there are no error bars or multi-seed runs anywhere. In a safety-critical framing, a single run is not enough to support a 'large margin' claim. The hyperparameters ω=100, γ=2, and the radius/height ranges are tuned on the STU validation set; the test set is clean, but the drop from validation to test (FPR@95 from 9.60 to 21.69) suggests some tuning sensitivity.\n\nThird, the Point Raise proxy assumption is a real limitation. Contracting road patches produces protrusions near the sensor; real anomalies can be overhead, off-road, or non-protruding. The cross-dataset SemanticKITTI result is encouraging but doesn't mechanistically close that gap. The paper acknowledges boundary cases in the limitations, but not this coverage issue.\n\nMinor: the 'reduces FPR@95 by 36%' claim in the introduction doesn't match any number I can trace in the tables, and no code is released.\n\nBottom line: this is a solid, useful paper with a reproducible-looking method and strong held-out results. It deserves a serious referee, but the authors should be required to run UEM on the STU test set, report variance or multiple seeds, and clarify the FPR claim. I'd take it to a reading group on anomaly segmentation.","headline":"Strong empirical LiDAR OOD detector with a clean relative-energy formulation; SOTA claim needs UEM on the test set and error bars before I'd trust the margin.","tokens_in":14926,"tokens_out":2619,"would_cite":true,"duration_ms":23601,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Relative energy — the log-odds of OOD versus ID — makes LiDAR point-wise outlier detection state-of-the-art on STU and SemanticKITTI.","keywords":["LiDAR","out-of-distribution detection","energy-based model","relative energy","point cloud segmentation","synthetic outlier generation","autonomous driving","anomaly segmentation"],"falsifier":"Measure REL's point-wise AUROC on the STU test set separately for naturalistic anomalies (e.g., flat road debris, bins) versus controlled, protruding object placements; if FPR@95 on naturalistic anomalies is substantially above the reported 21.69, the Point Raise training signal does not cover non-protruding hazards. Alternatively, replace Point Raise with a different synthetic OOD generator (e.g., random-scaled external object clusters) and check whether the ΔE boundary transfers; a large performance drop would show the specific synthetic geometry, not the relative energy score itself, is ess","tokens_in":13794,"feed_emoji":"🚗","tokens_out":5531,"duration_ms":48627,"temperature":0.7,"pith_summary":"The paper tries to establish that a simple change in score function — comparing out-of-distribution evidence against in-distribution evidence per point, rather than using absolute energy or maximum softmax probability — removes the overconfidence that makes LiDAR OOD detectors unreliable. It pairs this relative energy score with Point Raise, a geometry-aware synthesis that converts flat road patches into compact protrusions resembling real hazards, so training needs no external OOD datasets. On the STU benchmark the method reaches 97.85 AUROC and a 9.60 false-positive rate at 95% recall, roughly a quarter of the strongest baseline's error, while keeping closed-set panoptic quality nearly intact. The paper argues that this combination of a shift-invariant relative score and simple synthetic outliers gives autonomous driving systems a practical per-point OOD signal without sacrificing segmentation.","feed_headline":"New score catches LiDAR oddities with 97.85% AUROC","feed_subtitle":"Per-point log-odds of OOD vs known classes cuts false positives to 9.6% on STU","key_machinery":"The central object is the relative energy margin ΔE, the log ratio of summed negative-logit exponentials to summed positive-logit exponentials. Being a ratio of two aggregated energy terms, it is invariant to a constant shift in all logits, which directly addresses the calibration problem of raw energy scores. The companion machinery is Point Raise: its adaptive decay a = −log(d_min/d_max)/(γ(d_max−d_min)) maps a flat road patch into a compact, density-preserving protrusion by scaling each point by exp(−a·d_shift) and adding random heights, creating the negative samples that make ΔE's decision boundary learnable.","core_discovery":"The central discovery is that the relative energy margin ΔE = log( Σ_{i∈y⁻} exp(f_i) / Σ_{i∈y⁺} exp(f_i) ), which is exactly the log-odds of 'OOD versus ID' under the grouped softmax, is a trainable, shift-invariant OOD score. Trained with a logistic loss and a single imbalance weight, it outperforms both post-hoc scores (Max Logit, energy) and learned energy objectives (hinge, VOS, dual energy), cutting FPR@95 to 9.60 on the STU validation set versus 19.45 for the best energy baseline. The paper further shows that Point Raise — contracting a random road-point cluster radially toward the sensor with an adaptive decay and lifting it vertically — supplies sufficient auxiliary OOD data without","pith_inferences":["The synthetic protrusion bias may limit transfer to flat, non-raised anomalies such as road debris or overhead objects; a natural extension is to vary Point Raise with lateral or descending perturbations to test geometric coverage.","Because ΔE is exactly the softmax log-odds of a two-component mixture, the score can be interpreted as a log-likelihood ratio, suggesting principled conversion to probabilities or calibrated uncertainty estimates for downstream planning.","The reliance on road patches as the sole inlier surface assumes road geometry is always present; off-road or unstructured scenes (e.g., vegetation-heavy rural driving) might require using other planar inlier surfaces for synthesis.","A testable extension is to combine Point Raise with a small set of real hazard scans, using the synthetic augmentation to boost sample diversity and potentially close the gap to fully unsupervised OOD detection."],"forward_implications":["A per-point OOD score can be added as a lightweight auxiliary branch to a mask-transformer LiDAR segmentation model without substantial loss of panoptic quality (PQ within ~3.7 on STU and ~1.6 on SemanticKITTI).","Relative energy reduces FPR@95 to 9.60 versus 19.45 for the best prior energy objective, suggesting that ratio-based scoring is more robust than absolute-energy margin learning.","Point Raise removes the need for external OOD datasets and density post-processing; a single additional training sequence suffices for the STU fine-tuning, making the approach scalable to new deployments.","The method transfers across datasets: trained on STU-style geometry, it achieves 96.76 AUROC on SemanticKITTI outlier classes, outperforming likelihood-ratio and prototype-based baselines.","Even with a frozen backbone, the OOD projector alone yields 94.43 AUROC, so REL can serve as a cheap add-on to already-trained segmentation models."],"fun_headline_variants":["Relative energy score lifts LiDAR OOD detection","Log-odds scoring trims LiDAR false alarms","Energy gap trick sharpens 3D OOD spotting","Point Raise boosts LiDAR OOD robustness"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that Point Raise's radially contracted and lifted road patches adequately cover the geometric variety of real anomalies; the paper itself acknowledges residual failures on high-curvature surfaces and object boundaries, where sparse sampling makes reliable anomaly segmentation difficult.","fun_headline_variants_meta":{"raw":{"variants":["Relative energy score lifts LiDAR OOD detection","Log-odds scoring trims LiDAR false alarms","Energy gap trick sharpens 3D OOD spotting","Point Raise boosts LiDAR OOD robustness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000125,"raw_usage":{"total_tokens":963,"prompt_tokens":785,"completion_tokens":178,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":117}},"tokens_in":529,"tokens_out":178,"duration_ms":2471,"temperature":1.0,"reasoning_tokens":117,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T23:13:28.627861+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure REL's point-wise AUROC on the STU test set separately for naturalistic anomalies (e.g., flat road debris, bins) versus controlled, protruding object placements; if FPR@95 on naturalistic anomalies is substantially above the reported 21.69, the Point Raise training signal does not cover non-protruding hazards. Alternatively, replace Point Raise with a different synthetic OOD generator (e.g., random-scaled external object clusters) and check whether the ΔE boundary transfers; a large performance drop would show the specific synthetic geometry, not the relative energy score itself, is ess","supporting_citations":[],"review_version":1}