OpenNovelty.org

RECON: Robust symmetry discovery via Explicit Canonical Orientation Normalization

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
symmetry discoverycanonicalizationequivariance
Abstract:

Real world data often exhibits unknown, instance-specific symmetries that rarely exactly match a transformation group GG fixed a priori. Class-pose decompositions aim to create disentangled representations by factoring inputs into invariant features and a pose g∈Gg\in G defined relative to a training-dependent, \emph{arbitrary} canonical representation. We introduce \textsc{recon}, a class-pose agnostic \emph{canonical orientation normalization} that corrects arbitrary canonicals via a simple right-multiplication, yielding \emph{natural}, data-aligned canonicalizations. This enables (i) unsupervised discovery of instance-specific pose distributions, (ii) detection of out-of-distribution poses and (iii) a plug-and-play \emph{test-time canonicalization layer}. This layer can be attached on top of any pre-trained model to infuse group invariance, improving its performance without retraining. We demonstrate results on 2D image benchmarks and extend unsupervised instance-level pose discovery to 3D groups.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces RECON, a method for canonical orientation normalization that discovers instance-specific pose distributions without supervision. Within the taxonomy, it occupies the 'Canonical Orientation Normalization and Pose Discovery' leaf under 'Pose and Orientation Estimation via Symmetry'. Notably, this leaf contains only the original paper itself, with no sibling papers identified in the taxonomy. This suggests the specific combination of unsupervised instance-level pose discovery and canonical orientation correction via right-multiplication represents a relatively sparse research direction within the broader symmetry-based learning landscape.

The taxonomy reveals neighboring work in related but distinct areas. The sibling leaf 'Adversarial and Generative Pose Estimation for Symmetric Objects' addresses pose estimation through generative modeling rather than canonical normalization. Nearby branches include 'Symmetry-Based 3D Reconstruction' (focusing on shape learning from images) and 'Symmetry-Aware Representation Learning' (emphasizing equivariant features). The scope note for the paper's leaf explicitly excludes supervised pose estimation and methods not addressing instance-level variation, positioning RECON at the intersection of unsupervised learning and instance-specific adaptation—a boundary less populated than global symmetry-based approaches.

Among thirty candidates examined, the contribution-level analysis shows varied novelty profiles. The core RECON method for unsupervised pose distribution discovery examined ten candidates with zero refutable prior work, suggesting this specific formulation appears novel within the search scope. Similarly, the canonical orientation normalization via right-multiplication showed no clear refutations across ten candidates. However, the test-time canonicalization layer contribution examined ten candidates and found six that could refute it, indicating substantial prior work on adapting pre-trained models with symmetry-based transformations exists within the limited search.

Based on the top-thirty semantic matches examined, the paper's core methodological contributions around instance-specific pose discovery and right-multiplication normalization appear relatively novel, while the test-time adaptation component overlaps more substantially with existing work. The sparse population of its taxonomy leaf and absence of sibling papers suggest the specific problem formulation occupies an underexplored niche, though the limited search scope means potentially relevant work outside these candidates may exist.

Taxonomy

Core-task Taxonomy Papers
36
3
Claimed Contributions
30
Contribution Candidate Papers Compared
6
Refutable Paper

Research Landscape Overview

Core task: Unsupervised discovery of instance-specific symmetry distributions from unlabeled data. The field encompasses a diverse set of approaches that exploit symmetry as an inductive bias across multiple learning paradigms. At the highest level, the taxonomy organizes work into branches addressing 3D reconstruction and shape learning (e.g., Symmetric Shape Autoencoder[9], Dense Shape Correspondence[11]), pose and orientation estimation (e.g., Symmetric Keypoints[21], Reflection Symmetry Discovery[20]), symmetry-aware representation learning (e.g., Tri Invariance Contrastive[22], Symmetry Strikes Back[6]), and anomaly detection in specialized domains (e.g., Bridge Damage Detection[4], Railway Inspection Symmetry[12]). Additional branches cover domain adaptation (Two Branch Adaptation[10]), instance-specific augmentation (Instance Specific Augmentations[5]), clustering and pseudo-labeling (Cluster Matching Fisher[7], Speaker Verification Clustering[8]), neuro-symbolic verification (Verification Learning[1]), temporal and depth estimation (TSUDepth[3]), and theoretical foundations (Symmetry Meets AI[14], Permutation Symmetry Learning[33]). These branches reflect a shared recognition that symmetry constraints can guide learning even without explicit supervision, though they differ in whether symmetry is treated as a global prior, a local descriptor, or an instance-level distribution. Several active lines of work explore contrasting strategies for leveraging symmetry. One thread focuses on discovering canonical orientations or pose normalizations to enable consistent feature extraction, as seen in works like Symmetric Keypoints[21] and Reflection Symmetry Discovery[20]. Another emphasizes learning equivariant or invariant representations that respect known or discovered symmetries, exemplified by SymmetryLens[29] and Symmetry Strikes Back[6]. RECON[0] sits within the pose and orientation estimation branch, specifically targeting canonical orientation normalization and pose discovery. Compared to methods that assume fixed symmetry types (e.g., bilateral reflection in Symmetric Deformable Objects[2]), RECON[0] aims to infer instance-specific symmetry distributions in an unsupervised manner, aligning it with approaches that adapt to per-instance structure (Instance Specific Augmentations[5]) while maintaining a focus on pose discovery rather than augmentation or domain transfer. This positioning highlights an ongoing tension between imposing strong geometric priors and flexibly adapting to heterogeneous symmetry patterns across instances.

Claimed Contributions

RECON method for unsupervised instance-specific pose distribution discovery

The authors introduce RECON, a framework that discovers instance-specific symmetry distributions from unlabeled data by building upon class-pose decomposition methods. This enables unsupervised learning of probability distributions over transformation groups that describe how each instance appears in the data.

10 retrieved papers
Canonical orientation normalization via right-multiplication

The authors develop a canonical orientation normalization technique that corrects arbitrary canonical representations produced by class-pose methods. By estimating the Fréchet mean of observed relative poses and applying a simple right-multiplication, they obtain data-aligned canonicalizations that are robust, interpretable, and comparable across classes.

10 retrieved papers
Test-time canonicalization layer for pre-trained models

The authors provide a practical application in the form of a plug-and-play test-time canonicalization layer that can be added to frozen pre-trained models. This layer grants group invariance to existing models and improves their downstream performance without requiring architectural changes or retraining.

10 retrieved papers
Can Refute

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Within the taxonomy built over the current TopK core-task papers, the original paper is assigned to a leaf with no direct siblings and no cousin branches under the same grandparent topic. In this retrieved landscape, it appears structurally isolated, which is one partial signal of novelty, but still constrained by search coverage and taxonomy granularity.

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

RECON method for unsupervised instance-specific pose distribution discovery

The authors introduce RECON, a framework that discovers instance-specific symmetry distributions from unlabeled data by building upon class-pose decomposition methods. This enables unsupervised learning of probability distributions over transformation groups that describe how each instance appears in the data.

Contribution

Canonical orientation normalization via right-multiplication

The authors develop a canonical orientation normalization technique that corrects arbitrary canonical representations produced by class-pose methods. By estimating the Fréchet mean of observed relative poses and applying a simple right-multiplication, they obtain data-aligned canonicalizations that are robust, interpretable, and comparable across classes.

Contribution

Test-time canonicalization layer for pre-trained models

The authors provide a practical application in the form of a plug-and-play test-time canonicalization layer that can be added to frozen pre-trained models. This layer grants group invariance to existing models and improves their downstream performance without requiring architectural changes or retraining.

Table of Contents