acceptodds
Under review as a conference paper at ICLR 2027

Learning Quality-Conditioned Feature Displacements for LiDAR Out-of-Distribution Detection

Abstract

LiDAR out-of-distribution (OOD) detection is essential for reliable open-world perception, where unexpected observations beyond the training distribution may appear during real-world deployment. Existing LiDAR OOD methods learn in-distribution feature structures without explicitly modeling the systematic representation changes induced by varying observation quality. Consequently, quality-induced representation deviations of in-distribution observations may be either incorrectly suppressed or absorbed as uncontrolled intra-class variation, obscuring the distinction between sensing-induced variation and genuine OOD deviations. To address this issue, we propose QuaD, a framework that learns quality-conditioned feature displacements to explicitly characterize admissible representation variation under different observation qualities. Our method establishes canonical class references from reliable high-quality observations and predicts class-dependent, quality-conditioned feature displacements to explain sensing-induced representation shifts. The predicted displacements are constrained during training such that quality-induced shifts can be accommodated while displacement-corrected representations remain consistent with the canonical class structure. At inference, semantic, contrastive, and displacement evidence are interpreted according to their class- and quality-conditioned in-distribution distributions, enabling more reliable distinction between admissible sensing-induced variation and genuine OOD deviations. Extensive experiments on multiple LiDAR OOD benchmarks demonstrate that QuaD consistently achieves state-of-the-art results, with up to a 108.3% improvement in point-level AP on SemanticKITTI-OoD over the previous best method.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LiDAR Out-of-Distribution Detection, Feature Displacement Modeling, Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_P8EhOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes QuaD, a LiDAR out-of-distribution (OOD) detection method that distinguishes feature changes caused by sensing quality from those caused by unknown objects. Building on the dual-head LIDO framework, QuaD introduces two modules. QCRL learns class-specific canonical anchors from high-quality observations and predicts quality-conditioned feature displacements. QCAE converts semantic, contrastive, and displacement evidence into anomaly scores using class- and quality-conditioned quantile distributions. Experiments on nuScenes-OoD, SemanticKITTI-OoD, SemanticPOSS-OoD, and STU show consistent improvements over LIDO.

Critical strengths

The paper addresses an important LiDAR perception problem by explicitly modeling how distance and local sparsity affect normal feature representations, rather than simply adjusting anomaly thresholds, and the improvement from 6.79 to 20.94 AP with QCRL alone (Table 4) provides strong evidence for the value of this approach. Furthermore, the separate ablations of QCRL and QCAE, including the improvement from 20.94 to 26.66 AP when adding QCAE, along with experiments on conditioning variables and evidence fusion (Tables 10–11), help explain the method's contributions, while consistent improvements across multiple LiDAR OOD benchmarks demonstrate its potential.

Critical weaknesses
  1. The state-of-the-art claim is too broad because NDP-EE (CVPR 2026) reports much stronger STU validation results (74.24 AP and 1.43 FPR@95 versus QuaD's 40.25 AP and 19.26 FPR@95), although NDP uses a different backbone and synthetic OOD supervision, so the results do not directly invalidate QuaD's contribution under ID-only training; additionally, Section 5.2 reports ablations on the nuScenes-OoD test set without clearly explaining whether model selection was performed independently, although this alone does not establish test-set tuning.

  2. The effects of quality conditioning and the reliability of conditional quantile estimation are not sufficiently validated, as QCRL provides most of the improvement over LIDO (6.79 to 20.94 AP), while joint Class+Quality conditioning improves AP from 23.68 to 26.66 over unconditional QCAE (Table 10), which cannot be attributed to quality alone; moreover, the poor Class-only result (13.30 AP), high Quality-only FPR@95 (63.64), and lack of controlled sensing-degradation experiments leave open questions about interactions between conditioning variables, estimation stability, and whether distance and sparsity capture actual sensing quality rather than correlations with object geometry or class.

Questions for authors

Which training or validation split was used to fit the QCAE quantile model? Were the nuScenes-OoD test-set ablation results used for model selection, or was a separate development split used?

Reviewer_1vP3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets failures in LiDAR OOD segmentation caused by sensing degradation, where distant or sparse in-distribution points drift from their class prototype and get flagged as anomalies. Built on LIDO, QuaD learns a class-specific displacement conditioned on range and sparsity, so degraded features are compared against a quality-adjusted reference. It then scores anomaly evidence by how rare it is for each class at each quality level. Experiments on four benchmarks, including the real-world STU dataset, show clear gains over LIDO in AP and FPR@95 with negligible overhead.

Critical strengths
  1. The paper tackles a real, LiDAR-specific confound, namely range- and density-induced feature shift, which prior ID-based OOD methods ignore. Modeling it as a continuous displacement from a canonical class anchor is a clean alternative to discrete quality-specific prototypes.

  2. The results are strong. QuaD beats LIDO on all four benchmarks, including the real-world STU set, with big jumps in AP and FPR@95. It also barely costs anything, about 8.4K extra parameters and 3 ms per scan, and closed-set mIoU stays the same.

  3. The ablations are fairly complete. The authors compare against multi-prototype and distance-binned alternatives and also check the constraint space, conditioning factors, and fusion strategy, so it’s reasonably clear where the gains come from.

Critical weaknesses

[W1] No results are reported over multiple seeds, and all ablations are run on the nuScenes-OoD test set. AP on rare OOD points can be noisy, and some ablation differences are small. Mean and standard deviation, plus a note on how hyperparameters were chosen, would strengthen the results.

[W2] Heuristic observation-quality estimation. The quality score uses only range and point density, ignoring factors like occlusion, surface reflectivity, weather, and sensor beam count.

[W3] Table 4 reports no QCAE-only setting applied directly to LIDO.

Questions for authors

[Q1] Could the authors report results for QCAE applied directly to LIDO, without QCRL?

[Q2] Could the authors compare QuaD with recent LiDAR OOD methods such as NDP, REL, LiON, and DOODLE, at least on STU? If this is not feasible, could they explain why?

Reviewer_5bY1OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes QuaD, a LiDAR OOD detection framework built on LIDO that aims to distinguish sensing-induced representation variation from genuine OOD deviations. Using a per-point quality state derived from range and local voxel occupancy, QCRL models contrastive features as a canonical class anchor plus a class-dependent, quality-conditioned displacement. QCAE then interprets semantic, contrastive, and displacement evidence under class- and quality-conditioned in-distribution distributions. Experiments on three mixed real-synthetic benchmarks and the real-world STU dataset show consistent gains over LIDO, particularly in AP and FPR@95, with low computational overhead.

Critical strengths
  1. QuaD consistently improves OOD detection performance across multiple LiDAR benchmarks, particularly in AP and FPR@95, demonstrating its effectiveness under diverse sensing conditions.

  2. The paper provides extensive ablation studies and evaluations on both synthetic and real-world OOD benchmarks, supporting the effectiveness of the proposed components.

Critical weaknesses
  1. Lack of originality. The central motivation of QuaD—modeling sensing-induced feature variations caused by LiDAR range and point density—has been explored in prior work. DDFE [1] explicitly estimates range-dependent beam density and employs density-conditioned attention to modulate point- and voxel-wise features. Although DDFE focuses on domain generalization rather than OOD detection, its relevance to QuaD's quality-conditioned representation learning warrants further discussion. Furthermore, DOODLE [2] has already addressed false positives caused by non-uniform LiDAR measurements in the OOD detection setting, using depth-corrected density-aware post-processing to refine anomaly predictions. While DOODLE is cited in the paper, this closely related component is not discussed. The authors are encouraged to expand the related work discussion to clarify the methodological distinctions and advantages of QuaD over these existing density-aware approaches. In particular, a controlled comparison incorporating DOODLE-style density-aware refinement into the LIDO baseline would help establish the effectiveness of the proposed formulation.

  2. Insufficient analysis of segmentation performance. its closed-set segmentation performance remains considerably lower than the standard segmentation baseline, particularly on nuScenes-OoD (60.72 vs. 72.75 mIoU). Moreover, while Table 8 compares quality-conditioned constraints applied to shared, semantic, and contrastive feature spaces, it reports only OOD detection metrics. The corresponding mIoU results would help clarify how the choice of representation space affects the trade-off between OOD detection and semantic segmentation. Furthermore, the reliability of canonical anchors for underrepresented classes in sparse LiDAR observations remains insufficiently analyzed.

References

[1] Kim, Jaeyeul, et al. "Rethinking lidar domain generalization: Single source as multiple density domains." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.

[2] Oh, Changgyoon, et al. "DOODLE: Diffusion-based Out-of-Distribution Learning for Open-set LiDAR Semantic Segmentation." 2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2026.

Questions for authors
  1. How does QuaD's quality-conditioned feature displacement fundamentally differ from the density-aware feature modeling in DDFE [1] and the density-aware OOD refinement in DOODLE [2]? In particular, which specific algorithmic components distinguish QuaD from these prior approaches, and what empirical evidence justifies the necessity of the proposed displacement formulation?

  2. What are the individual contributions of sensing range and local voxel occupancy to OOD detection performance? Could the authors clarify whether both factors are necessary, for example, by reporting range-only and occupancy-only ablations? How sensitive is the performance to the weighting coefficient α\alpha in Eq. (3)?

  3. What explains the substantial closed-set segmentation performance gap on nuScenes-OoD (60.72 vs. 72.75 mIoU)? Given that Table 8 evaluates quality-conditioned constraints in shared, semantic, and contrastive feature spaces, could the authors also report the corresponding mIoU results to clarify the impact of these design choices on closed-set segmentation? Furthermore, how are reliable canonical anchors established for underrepresented classes with limited high-quality observations? A class-wise analysis of anchor reliability and displacement effectiveness would help clarify this issue.

  4. Provide the explicit mathematical definitions of LsheadL_{\mathrm{shead}} and LcheadL_{\mathrm{chead}} in Eq. (9), or specify the exact equations adopted from LIDO. These definitions are essential for understanding and reproducing the training objective.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.