acceptodds
Withdrawn from ICLR 2027

Adaptive Connectivity Is Not Correspondence: A Cross-Pose Study of Vision Graphs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision graph networks dynamically connect image regions through input-dependent edges, yet it remains unclear whether these learned connections capture true cross-view correspondence. We investigate this distinction using cross-pose face recognition as a controlled testbed. While achieving high frontal accuracy, the tested vision-graph models underperform a convolutional baseline on frontal-profile verification. Through controlled interventions, we find that grid anchoring emerges at an intermediate stage, while features at supplied correspondences retain a measurable advantage over identity-substituted controls. A geometric bound on cell-constrained sampling reach, however, does not imply a bound on recognition capacity. Under matched data and training budgets, pose augmentation alone outperforms adding the tested explicit correspondence alignment. Under a full 20-epoch schedule, pose augmentation improves extreme-pose genuine acceptance from 86.05% to 92.02 ± 0.99%. These recognition gains replicate across training datasets, whereas connectivity patterns are dataset-dependent, and the augmented models become less robust to pixel-noise corruptions. Together, our results show that adaptive connectivity, cross-view correspondence, and downstream recognition require distinct evidence and should be evaluated separately.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
vision graph neural networks, adaptive connectivity, cross-view correspondence, grid anchoring, representation learning, graph representation analysis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_54EYOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper uses cross-pose face recognition as a controlled testbed, building GraphFace on PViG-S to investigate whether the adaptive connectivity of vision graph networks captures true cross-view correspondence. Through controlled interventions, the authors find that grid anchoring emerges at an intermediate stage (14²), while features at supplied correspondences retain a measurable advantage over an identity-substituted control, and that the geometric bound on cell-constrained sampling reach does not imply a bound on recognition capacity. Under matched data and training budgets, pose augmentation alone outperforms adding the tested explicit correspondence alignment; under a full 20-epoch schedule it improves extreme-pose GAR from 86.05% to 92.02 ± 0.99%, a gain that replicates across training datasets, whereas connectivity patterns are dataset-dependent and the augmented models become less robust to pixel-noise corruptions.

Critical strengths

This article corrects an intuition and reminds everyone not to overinterpret the edges of graph networks.

Critical weaknesses
  1. The motivation is not well supported. The paper is framed as testing a "natural reading" that adaptive connectivity also captures cross-view correspondence, and says these mechanisms "are not always distinguished." I could not find any cited work that actually makes this claim or relies on it. ViG builds its graph within a single image, and the original paper mainly shows within-image semantic links, so I am not sure who holds the view being refuted here. The paper also does not show that this interpretation has led anyone to a wrong conclusion. On top of that, the one intervention that clearly helps is pose augmentation, which is already well known (Masi et al., 2016; Zhao et al., 2017), as the authors note. The question the paper starts from, namely why GraphFace falls behind the CNN at extreme yaw, is never really answered. So I find the overall insight limited.

  2. The evidence behind the main conclusions is thin, and important controls are missing. The claim that pose augmentation alone beats the tested alignment loss comes from one seed at five epochs (Table 4), so the seed variance of the −0.986 / +0.457 pp gaps is unmeasured (for reference, V's three-seed SD on CFP-FP is already 0.24 pp at 20 epochs), and only a single attractive-only cosine loss was tried. The grid-anchoring result is similarly weak: the JgridJ_{grid} and JcontentJ_{content} values come from a single MS1MV2 seed with no intervals, and although the 75° gap replicates on CASIA-WebFace, the yaw-averaged pattern does not. The controls that would settle these questions, such as a CNN trained on the same renders and graph models trained with GDC vs. average pooling, are listed by the authors themselves but were not run.

Questions for authors

Mentioned in Weaknesses.

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

This paper investigates whether dynamic vision graph networks implicitly learn true physical cross-view correspondence under severe viewpoint changes. Using cross-pose face recognition as a testbed, the authors perform systematic interventions, synthetic rendering controls, and geometric analysis to evaluate adaptive edge formation. The study demonstrates that dynamic graph connectivity collapses to fixed spatial grid coordinates under extreme pose rotations rather than tracking physical surface locations. Furthermore, the paper shows that cross-pose recognition gains stem primarily from pose data augmentation rather than explicit feature correspondence supervision.

Critical strengths
  1. Rigorously deconstructs a widely accepted assumption in dynamic vision graphs. By decoupling network connectivity, local feature correspondence, geometric sampling reach, and identity recognition performance into four distinct empirical and theoretical measurements, the authors provide essential scientific clarity. Demonstrating that adaptive edge formation fails to track physical surface locations and instead collapses to static grid anchoring under extreme viewpoint shifts exposes a major misconception driving current architectural designs.

  2. Provides a clean budget-matched isolation of data augmentation versus explicit feature alignment. Under strictly matched image count and optimization step budgets, forcing explicit feature alignment at corresponding facial locations via a cosine distance loss decreases face verification accuracy by 0.986% compared to pose augmentation alone. This proves that standard pose data augmentation fully accounts for cross-pose recognition gains, demonstrating that explicit correspondence supervision is unnecessary and can even be counterproductive for identity representation learning.

  3. Establishes theoretical boundaries on tracking error through a formal lower bound. Theorem 1 proves an exact lower bound on tracking error for cell-constrained query locations, providing clear geometric boundaries for dynamic sampling rules. The theoretical analysis distinguishes between the spatial reach of feature sampling coordinates bounded within local grid cells and the overall representation capacity of aggregated graph node features, mathematically clarifying why local sampling rules cannot track physical surface motion without unrestricted coordinate offsets.

  4. Validates claims through thorough synthetic rendering testbeds, cross-dataset replication, and numerical controls. To provide exact ground-truth surface correspondences and visibility masks across yaw angles from 0° to 75°, the authors fit three-dimensional morphable face models to CFP images, proving that grid anchoring dominates intermediate graph layers. They also test cross-dataset replication on MS1MV2 and CASIA-WebFace, showing that while downstream recognition gains from pose augmentation replicate consistently, specific learned edge structures do not, proving that graph connectivity patterns are dataset-dependent artifacts. This empirical protocol is further bolstered by analyzing subtle numerical sensitivities, showing how evaluation batch size variations alter top-k nearest neighbor graph topologies and verification accuracies by up to 0.467%.

Critical weaknesses
  1. The central limitation of this work is its evaluative nature, which critiques existing paradigms without proposing a novel architecture or replacement primitive. While the paper successfully proves that current dynamic vision graphs fail to learn physical correspondence and derives bounds on cell-constrained tracking error, it leaves the field at a diagnostic dead end by failing to introduce a modified sampling rule or architectural fix that resolves the grid anchoring phenomenon.

  2. The paper relies primarily on a single pyramid vision graph architecture rather than benchmarking a diverse array of dynamic graph paradigms across all primary experiments. While GraphFace serves as an effective testbed built upon PViG-S, alternative dynamic vision graph variants—such as windowed vision graph networks (WiGNet) and adaptive gating models (AdaptViG)—are utilized only as background references or diagnostic controls rather than fully trained external baselines across the main cross-pose evaluations.

  3. The conclusions regarding the inefficiency of explicit correspondence alignment are constrained by a narrow optimization scope that evaluates only a single attractive-only cosine distance loss over a five-epoch training schedule. This setup omits contrastive loss formulations, margin-based metric learning, or alternative spatial regularization functions, and a five-epoch training window may not capture the long-term convergence dynamics or spatial feature re-organization that occurs when fine-grained correspondence supervision is applied over a full 20-epoch training schedule.

  4. The paper highlights lower floating-point computational counts for vision graphs without adequately addressing their significantly higher real-world GPU inference latency. Although GraphFace requires only 1.49 GMACs compared to 6.33 GMACs for a standard ResNet-50 baseline, hardware benchmarking in the appendix reveals that GraphFace exhibits a batch-1 latency of 4.38 ms on an RTX 3090 Ti compared to 1.89 ms for ResNet-50, along with lower batch throughput. Presenting vision graphs as highly efficient alternatives purely based on MAC counts is misleading without integrating these latency tradeoffs directly into the main text discussion.

  5. Empirical evaluations using ImageNet-C and OODFace corruptions show that pose augmentation reduces mean verification accuracy across pixel-noise types by 2.86% on LFW and 1.51% on CFP-FP relative to the baseline graph model. While the paper correctly documents this degradation, it leaves the underlying mechanism unresolved rather than investigating whether top-k feature-distance graph construction is inherently sensitive to high-frequency pixel noise.

Questions for authors
  1. How do you propose addressing the grid anchoring phenomenon constructively in future dynamic vision graph designs without sacrificing spatial locality or feature aggregation efficiency?

  2. Can you provide empirical results using contrastive feature alignment or a full 20-epoch training schedule to verify whether explicit correspondence alignment remains inferior to pose augmentation under complete model convergence?

  3. Do alternative dynamic vision graph paradigms that use different neighbor selection mechanisms (such as WiGNet or AdaptViG) exhibit the same grid anchoring behavior under severe viewpoint rotations when trained end-to-end?

  4. What underlying mechanism accounts for the observed drop in pixel-noise robustness when vision graph models are trained with pose data augmentation?

Reviewer_25UQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This article examines vision graph neural networks (specifically GraphFace) in the context of facial verification across different poses. The authors argue that the dynamic connectivity of the graph (based on k-NN) fails to capture the true correspondence of 3D surfaces across different head rotations, demonstrating instead that the edges at 14x14 resolution stage remain anchored to the 2D image grid under extreme yaw conditions (75°), while local features evaluated based on reference 3DMM correspondences retain identity information. With Theorem 1, they present a geometric bound on coordinate sampling constrained to cells; they test several pilot interventions across short training schedules and conclude that standard 3DMM pose augmentation alone improves accuracy across different poses, although it degrades robustness to pixel-noise corruption. Although the article is transparent in reporting negative results and experimental confounding factors, its scientific contribution, technical novelty, theoretical depth, and experimental completeness fall well short of the standards required by ICLR. As detailed below, the architecture and effective intervention lack novelty, the formal theorems are trivial examples of geometry (and apply to a failed experiment that was never actually conducted rather than to the proposed model), the benchmarks and control experiments are either unfair or poorly formulated, and the manuscript reads like an incomplete forensic post-mortem analysis of an earlier draft, lacking even executable training code.

Critical strengths
  • Regarding the audit of experimental confounds, the discussion of evaluation pitfalls in Table 6 (and Appendix H) is methodologically sound and honest: it specifically demonstrates how identity-shuffled controls invalidate naive landmark diagnostics, how window-restricted baselines inflate the apparent gains of edge rewiring, and how fp32 batch-size changes (64 vs. 128) shift top-k accuracy by up to 0.467 pp.

  • The authors are also transparent about negative results, candidly reporting failed hypotheses, pre-registered gates (Table 21), and retracted explanations rather than selectively cherry-picking positive outcomes.

Critical weaknesses
  • Severe lack of architectural and methodological novelty. GraphFace merely recombines existing components (PViG-S backbone, PReLU, MobileFaceNet GDC readout, ArcFace loss), all standard, long-established face-recognition blocks. Its highly efficient motivation based on GMACs is misleading: appendix benchmarks show it is 2.3–3.2 x slower than IResNet-50 in wall-clock latency (GMACs omit neighbor-selection and feature-gathering costs), so it is both less accurate (7.53 pp GAR deficit at extreme yaw) and substantially slower. Every graph- or correspondence-specific intervention fails and the only improvement comes from synthetic 3DMM-rendered profile faces, a decade-old, graph-agnostic technique.

  • Trivial mathematical inflation and a broken, invalidated experiment. The main theorem and corollary are elementary high-school geometry as do the supporting proposition and lemmas. Moreover, the theorem does not apply to GraphFace, which has no dynamic sub-cell sampler; instead, it analyzes an auxiliary landmark-anchored lattice pilot that failed to receive landmarks during training due to a script bug (Appendix F.1) and was never re-run.

  • Near-tautological, dataset-dependent grid anchoring diagnosis. Both variants inject fixed 2D grid positional embeddings at the stem and every graph layer, so finding grid-anchored connectivity is largely expected; the authors admit positional ablations are needed but never run one. The headline grid-anchoring claim also fails its own pre-registered replication gate on CASIA-WebFace, where the yaw-averaged gap vanishes.

  • Flawed experimental controls. The conclusion that correspondence supervision underperforms rests on a single-seed, five-epoch pilot using an attractive-only cosine loss, a formulation known to collapse spatial feature variance, which directly explains the reported drop in correspondence specificity. Pose-augmented GraphFace is compared against an IResNet-50 never trained on the rendered pose pairs, violating experimental parity, and pose augmentation adds ~12.5% more images per step with no equal-image-budget control.

  • Incomplete manuscript and poor reproducibility. No runnable training code is provided, and key architecture details remain unspecified (Appendix A), undermining empirical reproducibility.

Questions for authors
  1. Why does the paper state that "Executable training code and checkpoint hashes are absent from the supplied manuscripts; these implementation choices cannot be inferred reliably from the equation"? Do the authors actually possess the executable training code and exact model definition for GraphFace, and if not, how can this work be reproduced?

  2. Given that the landmark-anchored lattice pilot never received landmarks during training due to the lmwarp loader bug (lines 1165–1177) and baseline GraphFace does not perform sub-cell sampling, what is the justification for keeping Section 4 and Theorem 1 as a centerpiece of the main paper rather than actually running the fixed experiment?

  3. Why was IResNet-50 not trained with the same 39,249 rendered pose pairs (even in a 5-epoch matched pilot) to verify whether pose augmentation specifically helps vision graphs or equally improves CNNs?

  4. Did you test a contrastive (InfoNCE) correspondence objective that penalizes similarity between non-corresponding spatial nodes to prevent the spatial feature collapse seen in Table 4 (𝐷_corr dropping from 0.156 to 0.034)?

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

Related papers

Discussion (0)

Sign in to comment.