acceptodds
Under review as a conference paper at ICLR 2027

SERVE: Semantic Evidence from Return Verification for LiDAR Segmentation

Abstract

An unfamiliar object can receive a confident semantic label when a LiDAR model only compares known classes. We propose Semantic Evidence from Return Verification (SERVE), which combines appearance support with class-conditioned range compatibility. A full-scan encoder supplies appearance support, while a separate predictor excludes the target angular cell and produces a distribution over log range for each class. Its density at the recorded range enters the supervised class decision alongside appearance support. The strongest joint class support supplies both the semantic label and the unknown score. Training uses normal semantic labels and recorded ranges. On STU, SERVE reaches 73.94% validation and 61.74% test average precision, compared with 21.36% and 15.09% for a matched appearance model, while improving normal semantic mIoU by 1.24 points. At a 1% normal false-positive rate, it recovers 81.6% of unknown points confidently assigned to known classes by the appearance model. Matched ablations assess the contributions of target exclusion, class-conditioned measurement support, and joint semantic training. These results support integrating learned range compatibility into semantic recognition.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LiDAR, anomaly segmentation, out-of-distribution detection, semantic segmentation, conditional density estimation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_fKuxOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

SERVE combines full-scan appearance features with class-conditioned log-range densities predicted from target-excluded neighborhood context for open-world LiDAR segmentation and anomaly detection. On STU, it improves test anomaly AP from 15.09% to 61.74% over a matched appearance baseline, while increasing normal-class mIoU by 1.24 points.

Critical strengths

Target-excluded, class-conditioned range verification is physically motivated and supported by ablations. Making the target cell available reduces validation AP from 73.94% to 39.62%, while removing class conditioning also degrades performance, supporting the importance of both design choices.

Critical weaknesses

The comparability of anomaly scores across different context states remains unclear. Empty-context points receive zero measurement energy, whereas supported points receive nonnegative energy. The paper does not report stratified detection results, leaving the impact of this score shift on pooled anomaly ranking uncertain.

Questions for authors

It would be helpful if the authors could report anomaly detection performance separately for points with and without available context. This would clarify whether SERVE retains its advantage over the appearance baseline after controlling for context availability or calibrating scores across context states.

1 comment
AuthorsOpenReview ↗

Response to Reviewer fKux

Thank you for suggesting the within-group evaluation. With neighborhood occupancy fixed, appearance alone and complete SERVE achieve 34.40% and 51.00% AP on identical returns with both scales occupied. With retrieval fixed, adding reference support increases full-validation AP from 46.57% to 50.81%. These experiments separate the within-group detection gain from changes in score ordering across context states.

During our review, we identified a systematic data-processing error affecting the submitted experiments. We take responsibility and withdraw the original results, including 73.94% validation AP, 61.74% hidden-test AP, the semantic accuracy gain, and the target-exclusion ablation you highlighted. The original predictions and checkpoints needed to locate the exact fault are unavailable. We rebuilt pointwise evaluation using public STU labels and original return indices. The anonymous supplement provides reproducibility checks and full-validation evaluation code.

1. Could context-dependent score scales explain the submitted gains?

In the submitted design, empty neighborhoods contributed zero range energy. An offset between supported and unsupported returns could change pooled AP without changing either group's internal ranking.

We reconstructed the angular neighborhood and target exclusion on 128 previously examined scans. Among supported returns (12,353,019 normals and 5,762 anomalies), frozen appearance alone achieved 0.621% AP and 30.45% FPR95. Adding a previously fitted range head with per-class calibration and weight 0.1 gave 0.898% AP and 25.22% FPR95.

Only 35 returns lacked neighbors, all normal. Their range energy was zero, their scores equaled appearance scores, and none crossed either readout's calibrated cutoff.

A separate matched range-head comparison yielded 0.236% AP because it used a reinitialized head selected at update zero, raw energy at weight 1, and all eligible returns. The submitted jointly trained checkpoint remains unverified.

2. Does the revised detector improve ranking within context states?

The revised model evaluates the probability assigned by neighboring semantic predictions to four retrieved normal references. When both neighborhood scales are empty, the reference distribution equals the range prior, making the reference discrepancy zero. Other anomaly signals remain available.

The table compares appearance → complete SERVE on identical native validation returns. AP and FPR95 are computed within each group; FPR and recall use each readout's fixed normal-sequence-201 cutoff.

Occupied scalesNormal / anomaly returnsAP (%)FPR95 (%)FPR at cutoff (%)Recall at cutoff (%)
Zero3,881 / 0——15.41 → 22.75—
One176,109 / 18520.06 → 25.2633.83 → 19.145.60 → 7.9251.35 → 65.41
Both193,612,480 / 87,31434.40 → 51.007.18 → 3.530.955 → 0.50480.34 → 85.14

The two-scale group contains 99.79% of anomalies. Its +16.60-point AP gain closely matches the pooled +16.26-point gain. A between-group offset cannot produce this within-group improvement. AP is nonlinear, however, so these comparisons do not quantify the exact contribution of cross-group ordering.

The fixed thresholds reveal a tradeoff: among returns with one occupied scale, SERVE detects 26 additional anomalies but flags 4,089 additional normal returns (Appendix D, Tables 9–10).

At native resolution, no anomalies have both scales empty. Thinning to 32 rows produces 108 such anomalies and 30,820 normal returns. At the unchanged cutoff, SERVE detects 83/108 anomalies (76.85%) and flags 10,599/30,820 normals (34.39%). This evaluates the complete detector under thinning (Appendix G).

3. Isolating reference verification

Table 2 fixes appearance, geometry, retrieved identities, distances, and spatial aggregation. Adding only reference support raises AP from 46.57% to 50.81% (+4.24 points; 95% paired sequence interval [2.91, 5.47]). With feature residual already included, the additional gain is +0.41 points [−0.24, 0.68].

Complete SERVE achieves 50.51% AP on full STU validation, above the previous published best we found for LiDAR-only methods fitted without real or synthetic anomaly examples (LIDO, 27.53%; Table 1). Final model selection used 320 of 1,960 scans across all 19 sequences. An official hidden-test result remains pending.

Reviewer_JCu5OpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SERVE to enable models to recognize known objects while also detecting objects outside the known classes. The method combines point-cloud appearance features with range predictions based on neighboring points, which are then compared with the actual measurements. It is trained using only normal semantic labels and recorded ranges. Its effectiveness is evaluated through matched experiments on STU and ablation studies of target exclusion, joint training, and scoring strategies.

Critical strengths

The ablation studies provide convincing evidence for the effectiveness of the key designs. Two controlled experiments support the contributions of joint semantic training and target exclusion, respectively, demonstrating that the performance improvements are directly related to the proposed mechanisms.

Critical weaknesses

The model's performance when no valid neighboring points are available, as well as the additional benefits of class-conditioned range prediction, requires further clarification.

Questions for authors
  1. Clarification: How does the model perform when no valid neighboring points are available? In this case, the model relies solely on appearance information. What proportions of known and unknown points fall into this category? What are the corresponding false-positive and missed-detection rates? These results would help assess whether the method is prone to missing small, isolated objects.
  2. Suggested additional experiment: How much additional benefit does class-conditioned range prediction provide? Have the authors compared SERVE with a model that predicts range from neighboring points without conditioning on object classes and then combines range predictions with appearance information? We suggest training such a baseline under the same data and training settings to quantify the additional contribution of class-conditioned range prediction.
1 comment
AuthorsOpenReview ↗

Response to Reviewer JCu5

Thank you for suggesting the shared range predictor. We conducted the matched comparison and examined empty-context failures, including sparse obstacles.

During our review, we identified a systematic data-processing error affecting the submitted experiments. We take responsibility and withdraw the original results, including 73.94% validation AP, 61.74% hidden-test AP, the semantic accuracy gain, and the joint-training and target-exclusion ablations. The surviving records cannot identify the exact fault. We rebuilt evaluation from public STU labels, matching each prediction to its original return under official eligibility rules.

1. What happens without neighboring context?

The revised detector uses neighboring class predictions to assess four retrieved normal references. When both neighborhood scales are empty, their predicted probabilities revert to the range prior, making the reference discrepancy zero. Other anomaly signals remain active.

At the cutoff calibrated on normal sequence 201:

InputEmpty normalsEmpty anomaliesFalse alarmsMisses
128 rows3,8810883/3,881 (22.75%)Undefined
32 rows30,82010810,599/30,820 (34.39%)25/108 (23.15%)

Complete context loss therefore explains none of the native-resolution anomaly misses. Sparse obstacles remain challenging: 943/2,800 scan-instances with fewer than 100 anomaly returns (33.68%) failed to reach 50% detection coverage, compared with 14/192 (7.29%) among instances with at least 100 returns. Relative to distance alone, 174 instances gained coverage and 16 lost it. Both methods missed the criterion on 941 instances (median five returns). Return count reflects sampling density, not necessarily physical size (Appendices D and G).

In a reconstruction of the submitted angular neighborhood on 128 previously examined scans, only 35/12,353,054 normals and 0/5,762 anomalies had no neighbors. None of those 35 normals triggered an alarm under the calibrated range readout.

2. What does class-conditioned range prediction add?

We compared the submitted class-conditioned head with a shared range predictor using identical frozen features, target exclusion, fitting scans, minibatch order, and 512-update budget. Normal-development mIoU selected the initial weights of the reinitialized conditioned head and update 512 of the shared head.

Using raw range energy at weight 1 on the same 128 scans:

ReadoutAPFPR95
Appearance0.621%30.45%
+ Class-conditioned range0.236%34.40%
+ Shared range0.285%53.00%

Neither head improved AP over appearance. Conditioning reduced FPR95 relative to the shared head but had lower AP in 16/19 sequences. Fully converged comparisons remain open.

The earlier context-state result of 0.898% AP used a previously fitted head, per-class energy calibration, weight 0.1, and only returns with neighbors. The 0.236% result uses a reinitialized head at update zero, raw energy at weight 1, and all eligible returns. These results have different settings and evaluated populations.

3. What does revised reference verification contribute?

Both methods test whether the surroundings support the center return. The revision predicts support for four retrieved normal references instead of predicting range under candidate classes.

With identical architecture, fitting queries, and targets, center-class conditioning plus range achieved 41.19% AP, versus 50.94% with neighboring class mixtures plus range. This ordering held across three paired fits (Table 3, Appendix H).

At fixed retrieval, adding reference support increased full-validation AP from 46.57% to 50.81% (+4.24 points; 95% paired sequence interval [2.91, 5.47]). Its additional gain alongside feature residual was smaller: +0.41 points [−0.24, 0.68] (Table 2).

The selected detector reached 50.51% AP on full STU validation, exceeding the previous published best we found for LiDAR-only methods fitted without real or synthetic anomaly examples (LIDO, 27.53%; Table 1). Model selection used 320/1,960 validation scans. The official hidden-test score remains pending.

Reviewer_K674OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SERVE, a LiDAR segmentation method that combines semantic appearance information with class-conditioned range prediction to better detect unknown objects that may be confidently classified as known classes. An appearance encoder processes the full LiDAR scan, while a separate context encoder predicts the log-range distribution for each candidate class without seeing the target cell. The method combines appearance support with the likelihood of the observed range to predict semantic labels and detect unknown points. The range predictor is trained jointly with semantic classification, while predictive negative log-likelihood (NLL) and appearance compactness losses help maintain reliable class support. On STU, SERVE improves validation AP from 21.36 to 73.94, hidden-test AP from 15.09 to 61.74, and normal mIoU from 61.92 to 63.16 compared with a matched appearance-only baseline.

Critical strengths
  1. The paper introduces a clear and effective approach that predicts class-conditioned range distributions from neighboring LiDAR cells without accessing the target cell, providing physical measurement evidence beyond semantic confidence, and the large AP improvements and recovery of confidently misclassified unknown points at 1% FPR demonstrate its practical value.

  2. The carefully designed ablations, including target-cell access, separate-objective training, common-density prediction, and independent evidence minimization, help isolate the contributions of the proposed components, while results across three seeds, hidden-test evaluation, predictive uncertainty analysis, and computational measurements strengthen the experimental evidence.

Critical weaknesses
  1. The generalization to different LiDAR sensors remains unclear because all major experiments use STU, while SERVE relies on fixed 0.5° angular cells and 16 neighboring context positions at radii of one and three cells; for example, a 32-beam LiDAR with an approximately 41° vertical field of view has an average beam spacing of about 1.3°, which may reduce vertical context coverage, although actual spacing and grid alignment vary and horizontal neighbors may remain available, so the impact on empty-context fallback and OOD performance must be measured rather than assumed.

  2. The effect of empty-context fallback and score calibration is not sufficiently evaluated because points with no valid context (kappa_i=0) use appearance-only scores, creating a different score scale from points with measurement evidence, while the paper does not report context-specific AP, FPR95, or score distributions; additionally, the 1% FPR analysis sets thresholds using the evaluation normal population, leaving uncertainty about performance when thresholds are calibrated on an independent normal dataset and transferred to another split.

Questions for authors
  1. Could the authors evaluate SERVE using beam subsampling on STU or another LiDAR benchmark with a different sensor resolution, reporting valid context coverage, the proportion of (kappa_i=0) points, range prediction error, and AP/FPR95? A sensitivity analysis of the 0.5° cell size and neighbor radii would also be helpful.

  2. If the 1% FPR threshold is selected using an independent normal calibration set and then kept fixed, how well do FPR and confident-unknown recall generalize to the validation and test sets?

1 comment
AuthorsOpenReview ↗

Response to Reviewer K674

Thank you for the 32-beam example. At the same alarm cutoff, thinning reduced anomaly detections from 74,459 to 58,050. Returns retaining neighbors at both scales accounted for 16,253 of the 16,409 net lost detections (99.05%). Even within that group, AP fell from 52.04% to 34.30%. On native STU validation, the cutoff calibrated on separate normal scans detected 59.66% of anomalies confidently assigned to known classes.

During our review, we found a systematic data-processing error affecting the submitted experiments. We take responsibility and withdraw the original claims, including 73.94% validation AP, 61.74% hidden-test AP, 81.6% confident-unknown recall, the semantic gain, and the original ablation and three-seed results. The surviving records cannot identify the exact faulty operation. We rebuilt pointwise evaluation using public STU scans and labels, matched scores to original returns, and applied official eligibility rules. Independent ranking calculations agree with the official evaluator for all three main checkpoints (Appendix B).

1. What does lower vertical resolution cost?

We partitioned the 128 sensor rows into four interleaved 32-row inputs and reran the complete detector. Every eligible return across 1,960 scans was scored once at each density, using the same alarm threshold.

InputAnomalies with both scales occupiedAPFPR95Recall at cutoffFPR at cutoff
128 rows87,314/87,49950.51%3.55%85.10%0.512%
32 rows84,435/87,49932.42%18.50%66.34%0.635%

Fully empty neighborhoods increased from 0.0020% to 0.0159% of normal returns and from zero to 0.123% of anomalies. A scale counts as occupied when at least one neighboring cell contains a return.

Among 84,435 anomalies with both scales occupied before and after thinning, detections dropped from 72,212 to 55,959. Thus, most lost detections arise despite retained neighborhood occupancy. Thinning also changes encoded features and retrieved references, so this experiment measures their combined effect. We used Cartesian cell widths of 0.75 and 2.25 m; sensitivity to each width and transfer to another sensor remain open (Appendix G, Table 14).

Both the submitted and revised methods assess whether surrounding observations support the center return. The submitted predictor estimated range under candidate classes. In its frozen reconstruction, alternate 0.5-degree elevation bins reduced mean absolute log-range error from 0.0721 to 0.0692 for one offset and from 0.0706 to 0.0689 for another, among normal returns supported at both densities. Detection AP nevertheless fell for both offsets. The revision predicts support for four retrieved normal references instead.

2. Does a normal-data threshold transfer?

We calibrated each readout at the 99th percentile of 2,304,711 verified normal returns from 32 scans of sequence 201. This sequence is disjoint from anomaly validation, although it was also used for upstream decoder selection.

At the fixed SERVE cutoff of 3.38604, native validation yielded:

  • 74,459 detected anomalies and 991,452 false alarms.
  • 85.10% recall, 0.512% FPR, and 6.99% precision.
  • Normal FPRs of 22.75%, 7.92%, and 0.504% for zero, one, and two occupied scales, respectively.

At the stricter 0.1%-calibration-FPR cutoff, these groupwise FPRs became 6.44%, 1.41%, and 0.035%. Appendix D (Tables 9–10) also reports within-group AP and FPR95.

Among 16,747 anomalies with known-class probability at least 0.9, reference verification increased recall from 54.21% (distance) to 55.39%, a net gain of 198 detections. With feature residual already included, recall changed from 59.99% to 59.69% (49 fewer). Full SERVE reached 59.66%, versus 47.25% for appearance. Relative to distance, full SERVE recovered 1,146 anomalies and lost 234, while introducing 10,428 additional normal alarms on balance within this confidence group (Appendix E, Table 12). Hidden-test FPR and confident-unknown recall remain unverified.

3. Contribution and remaining limitations

With appearance, geometry, retrieved references, and spatial aggregation held fixed, adding reference support raised full-validation AP from 46.57% to 50.81% (+4.24 points; 95% paired sequence interval [2.91, 5.47]). Its additional contribution alongside feature residual was smaller: +0.41 points [−0.24, 0.68] (Table 2).

The selected complete system achieved 50.51% AP, exceeding the previously published best we found for LiDAR-only methods fitted without real or synthetic anomaly examples (LIDO, 27.53%; Table 1). Model selection used 320 of 1,960 validation scans; AP on the remaining 1,640 was 50.29%, although 275 had been inspected during earlier development. An official hidden-test result for this checkpoint remains pending.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.