acceptodds
Desk rejected at ICLR 2027

RIFT: Registration-Induced Feature Tuning for Few-Shot 3D Anomaly Localization

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Few-shot 3D anomaly localization aims to identify subtle geometric defects from only a few normal reference scans. Existing registration-based methods primarily use alignment to reduce spatial ambiguity during feature matching or representation learning, leaving the structure exposed by registered normal references underexplored after the representation has been fixed. We propose RIFT, a lightweight Registration-Induced Feature Tuning framework that turns these aligned normal references from passive matching memories into supervision for post-hoc target-category specialization. RIFT exploits two complementary cues revealed by registration: canonical coordinates provide spatial identity for a coordinate-conditioned residual update, while same-slot patches across aligned templates provide cross-template structural correspondence supervision. The residual formulation preserves the pretrained representation while adapting its local metric without retraining the encoder or requiring anomalous samples and defect annotations. On Real3D-AD, RIFT improves P-AUROC from 94.21% to 96.43% and P-AUPR from 30.29% to 45.95%. On Anomaly-ShapeNet V2, it improves P-AUROC from 94.65% to 95.10% and P-AUPR from 44.40% to 46.19%. Controlled ablations further show that learned spatial conditioning and correct same-slot correspondence are both important, with the latter providing a clear advantage over generic or shuffled correspondence supervision.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
3D Anomaly Detection;Point Cloud Anomaly Localization;Self-Supervised Representation Learning

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_xHU8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RIFT, a lightweight post‑hoc feature adaptation method for few‑shot 3D anomaly localization using a frozen registration‑aware encoder and aligned normal templates. Registration assigns patches to predefined canonical slots and coordinates, treating cross‑template same‑slot patches as positives and different‑slot patches as negatives. RIFT combines frozen normalized patch features with Fourier‑encoded coordinates via layer‑normalized small MLP to generate gated residual feature updates, trained with InfoNCE‑style alignment and cosine compactness losses using only target‑category normal data. Adapted features serve for point‑level localization while the original object‑level detector is preserved. Built on COSSAD/RIConv++, RIFT improves Real3D‑AD P‑AUROC/P‑AUPR and Anomaly‑ShapeNet V2 P‑AUROC/P‑AUPR. Its adapter has 36,855 parameters, with ~7.4 s per‑category training and 0.12 ms overhead for 1000 inference descriptors. Real3D‑AD ablations cover fixed coordinates, coordinate‑only adaptation, shuffled correspondences, generic feature‑neighbor supervision, compactness‑loss removal and PointMAE backbone transfer.

Critical strengths
  1. Low-cost target-category specialization with meaningful Real3D localization gains. RIFT improves point AUPR by 15.66 points on Real3D-AD while freezing the backbone, using no anomalies, and adding very few parameters and little inference latency.
  2. Reasonably controlled internal evidence for correspondence supervision. The Frozen, fixed-coordinate, coordinate-only, shuffled-correspondence, and generic-neighbor comparisons suggest that the result is not solely caused by adding an adapter or exposing coordinates. Correct same-slot supervision appears beneficial on Real3D-AD.
Critical weaknesses
  1. The technical contribution is too incremental in its current form. COSSAD already learns from registered patch pairs; RIFT moves this supervision to post-hoc target-category adaptation and combines it with Fourier coordinates, a gated residual MLP, InfoNCE, and a compactness term. Each mechanism is standard, and the paper does not establish a new learning principle beyond applying registered correspondence supervision to a frozen adapter.
  2. The evidence for generality and for the central correspondence assumption is inadequate. RIFT assumes canonical slot identity implies structural equivalence, but there is no registration-error or correspondence-corruption analysis. The second benchmark improves by only 0.45 AUROC and 1.79 AUPR, 12/40 categories do not improve, no multi-seed uncertainty is reported, and the main ablations are limited to Real3D. The near-perfect slot-retrieval statistic may also be partly solved by the coordinate input itself.
Questions for authors
  1. How does RIFT behave under controlled rigid‑pose error, non‑rigid misregistration, partial overlap, symmetry‑induced swaps, and random corruption of a known fraction of slot correspondences? Could the authors report performance relative to Frozen as error increases, and identify conditions under which adaptation becomes harmful?
  2. Could the authors supply three‑run mean and standard‑deviation results across both datasets, and reproduce the shuffled‑correspondence, coordinate‑only, fixed‑coordinate, and generic‑neighbor ablations on Anomaly‑ShapeNet V2? I am curious whether those smaller performance gains are statistically reliable.
  3. The paper notes that the object‑level detector remains unchanged, yet the reported object‑level AUROC for RIFT diverges from published COSSAD numbers. Could the authors include a Frozen object‑level row evaluated within this identical pipeline and explain all evaluator‑related discrepancies.
  4. What exact normal‑only retrieval statistic drives checkpoint selection? Given the roughly 99.7–99.9 % held‑out slot‑retrieval performance, how much of this statistic can be solved purely from coordinate inputs? Could the authors provide a coordinate‑only retrieval control alongside full pair‑sampling and batch configuration details.
  5. How does RIFT compare against simpler post‑hoc target‑adaptation approaches constrained to matching trainable‑parameter and runtime budgets, for example linear metric learning, feature whitening, per‑slot prototypes, or adapters trained without Fourier coordinates? Could the authors include this comparison to establish whether the proposed gated‑residual architecture delivers meaningful benefits?
  6. I am curious whether every target category demands a separately stored and trained adapter. How do total storage overhead, calibration latency, and performance scale with growing numbers of categories and templates? Could the authors report any catastrophic degradation or cross‑category interference observed when training one single shared adapter across categories.
1 comment
Reviewer_xHU8OpenReview ↗

Official Comment by Reviewer xHU8

I appreciate the authors' detailed response and careful answers to my questions, which clarified many of my doubts. However, I believe my initial overall rating still accurately reflects my assessment of the paper, and therefore I would like to keep my original score.

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

This paper proposes post-hoc target-category adaptation for few-shot 3D anomaly localization on top of a frozen registration-aware detector. Normal reference scans are aligned into a shared canonical frame. Registered coordinates condition a gated residual update of pretrained patch descriptors, while patches assigned to the same canonical slot across different normal templates provide positive pairs for a contrastive and compactness objective. The encoder and registration pipeline remain frozen, and the original object-level detector is unchanged. In paired comparisons with Frozen COSSAD, RIFT improves point-level AUROC and AUPR from 94.21% and 30.29% to 96.43% and 45.95% on Real3D-AD, and from 94.65% and 44.40% to 95.10% and 46.19% on Anomaly-ShapeNet V2. The paper also reports controlled ablations, qualitative examples, efficiency measurements, a generic adaptation control, and a cross-backbone experiment.

Critical strengths

The paper provides well-matched evidence for its proposed mechanism. Frozen COSSAD and RIFT share the evaluation pipeline, while Table 4 separates a fixed spatial cue, learned coordinate conditioning, and correct same-slot supervision. On Real3D-AD, correct rather than shuffled correspondence increases P-AUPR from 37.57% to 45.95%. Appendix A.2 provides an additional nearest-neighbor adaptation control. These experiments make the interpretation more convincing than a performance comparison alone. The localization gains have practical value at low adaptation cost. Real3D-AD P-AUPR increases by 15.66 percentage points over the paired frozen baseline, with smaller positive aggregate improvements on Anomaly-ShapeNet V2. The target-category adaptation uses normal references only, leaves the encoder unchanged, and adds 36,855 parameters. Section 4.4 reports approximately 7.4 seconds of fitting per category and 0.12 milliseconds per 1,000 descriptors for the adapter.

Critical weaknesses

Robustness to imperfect registration is not established. The main training signal assumes that canonical slots correspond to the same structural role across normal scans. Although the authors acknowledge this dependence, the paper does not quantify sensitivity to alignment error or incorrect slot assignment. This matters because correspondence quality is central to the mechanism rather than an incidental implementation detail. The stability and generality of the improvements remain insufficiently characterized. The submission does not show repeated-run uncertainty or variation across adaptation seeds and reference choices. This is especially important for Anomaly-ShapeNet V2, where the aggregate P-AUROC increase is 0.45 percentage points and 28 of 40 categories improve. The most informative correspondence ablation is confined to Real3D-AD. The PointMAE transfer experiment is positive but starts from a substantially weaker frozen baseline.

Questions for authors

How sensitive is RIFT to registration inaccuracies? A controlled perturbation study or a breakdown by alignment-quality measure would help establish the reliability of same-slot positive pairs under imperfect registration. How stable are the paired localization improvements across independent adapter initializations and, where feasible, different normal-reference selections? Per-category P-AUPR variability or uncertainty estimates would be especially informative for Anomaly-ShapeNet V2. Does the correct-versus-shuffled correspondence ablation show the same pattern on Anomaly-ShapeNet V2? A matched replication would strengthen the claim that the gain arises from the proposed structural supervision rather than dataset-specific factors.

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

This paper proposes RIFT, a lightweight post-hoc feature adaptation framework for few-shot 3D anomaly localization. The method leverages registered normal reference point clouds to provide coordinate-conditioned feature adaptation and cross-template same-slot correspondence supervision. A residual adapter is optimized using contrastive alignment and compactness losses while keeping the pretrained encoder and registration procedure frozen. At inference, the adapted features are used for point-level anomaly localization, while object-level scoring remains unchanged. Experiments on Real3D-AD and Anomaly-ShapeNet V2 demonstrate improvements in localization performance.

Critical strengths

S1. The idea of exploiting registration-induced spatial correspondence as supervision for post-hoc feature adaptation provides an interesting perspective on few-shot 3D anomaly localization. The distinction between geometric similarity and structural identity is well motivated.

S2. The proposed framework introduces only a lightweight residual adapter without retraining the pretrained encoder or requiring target-category anomaly annotations. With 36,855 trainable parameters and approximately 7.4 seconds of adaptation per category, the method is computationally efficient.

S3. The paper includes well-designed controlled comparisons and ablation studies. In particular, the comparisons between coordinate-only adaptation, shuffled correspondence, and correct same-slot correspondence, along with the GNA-NN control, provide convincing evidence supporting the proposed design.

Critical weaknesses

W1. The performance improvements are uneven across the two evaluated benchmarks and different object categories. While RIFT achieves substantial gains on Real3D-AD, the improvements on Anomaly-ShapeNet V2 are relatively modest, with P-AUROC increasing from 94.65% to 95.10% and P-AUPR from 44.40% to 46.19%. Moreover, several categories exhibit performance degradation, such as bowl0, helmet2, and vase1. Further analysis of these cases would help clarify the conditions under which the proposed adaptation is effective.

W2. The proposed method relies on accurate registration and reliable same-slot correspondence across normal reference templates. Although the authors acknowledge that residual misalignment may affect the adaptation signal, this vulnerability is not quantitatively investigated. Controlled experiments with different levels of registration errors would help establish the robustness and practical applicability of the proposed approach.

W3. The evaluation mainly considers settings with four normal reference scans per category, without systematically investigating the effect of reference-set size. Since RIFT relies on cross-template correspondences for feature adaptation, its effectiveness with fewer normal references remains unclear. Experiments under different few-shot settings, particularly with one or two normal scans, would strengthen the practical contribution of the proposed framework.

Questions for authors

Could the proposed method further improve the performance of zero-shot approaches in few-shot setting?

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

Related papers

Discussion (0)

Sign in to comment.