acceptodds
Under review as a conference paper at ICLR 2027

Quaternion Structural-Spectral Aggregation for Training-Free Fine-Grained Leaf Image Retrieval

Abstract

Training-free fine-grained leaf image retrieval is attractive for continually expanding cultivar collections because it avoids target-specific annotation and fine-tuning. Existing approaches mainly summarize frozen feature maps using real-valued statistics or local pattern transforms, leaving both intra-patch two-dimensional organization and contour-wise cyclic structure underexplored. We propose Quaternion Structural-Spectral Aggregation (QSSA), a training-free descriptor built on multi-scale boundary-aligned patches and frozen CNN features. For each feature channel, QSSA constructs a quadrant-layout quaternion and a symmetry-discrepancy quaternion to encode absolute spatial layout and paired-region relations. The resulting contour-ordered quaternion sequences are aggregated using a left-sided discrete quaternion Fourier transform (DQFT). Quaternion magnitudes provide invariance to the arbitrary contour starting point, while low-frequency truncation suppresses high-frequency variations associated with local boundary perturbations. Experiments on two publicly available fine-grained leaf image retrieval datasets, SoyCultivar200 and PeanCultivar120, demonstrate that QSSA achieves state-of-the-art or competitive performance without target-specific training.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Fine-Grained Leaf Image Retrieval, Zero-Shot Image Retrieval, Local Deep-Structural Projection, Boundary Dynamic Patch, Geometric Projection, Fourier Feature Aggregation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9RHLOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

QSSA addresses training-free fine-grained leaf image retrieval, where the goal is to retrieve images of the same cultivar without fine-tuning a model on the target dataset. It extracts multi-scale patches along leaf boundaries, encodes frozen CNN feature maps using quadrant-layout and symmetry-discrepancy quaternions, and aggregates contour-ordered descriptors with a left-sided DQFT. The method uses quaternion magnitudes and frequency truncation to obtain a compact representation that is invariant to the choice of contour starting point.

Experiments on SoyCultivar200 and PeanCultivar120 show competitive or leading retrieval performance against the reported baselines. Ablations investigate sampling density, geometric scales, frequency retention, backbones, quaternion encodings, and aggregation alternatives. The main unresolved question is whether the measured gains establish a generalizable methodological advance rather than a task-specific combination of spatial encoding and spectral aggregation.

Critical strengths
  • The pipeline has a clear motivation: preserve local spatial organization while incorporating contour-order information.
  • The two quaternion branches represent different types of information: absolute regional responses and paired-region discrepancies.
  • The DQFT shift-invariance derivation is mathematically plausible under the paper's stated left-sided transform convention and fixed contour direction.
  • The method does not require training a new retrieval network for each cultivar collection.
  • The authors provide several component ablations rather than relying solely on comparisons against complete systems.
Critical weaknesses
  1. Table 5 reports 66.70% for Quat.-4ary and 63.11% for Real-4ary on PeanCultivar120, a 3.59-point difference. This is evidence in favor of the proposed representation, but it is not enough to establish the underlying mechanism. The paper needs to explain precisely how the real-valued control is constructed, how its normalization and dimensionality are matched, and which operations differ from the quaternion version. A quaternion transform with a fixed imaginary axis and magnitude-based output may be beneficial because it couples components during spectral processing, but the paper does not yet isolate this explanation from other implementation differences.
  2. The quadrant-layout branch and symmetry-discrepancy branch are plausible complementary encodings. However, the gains from combining them are relatively small compared with the individual branches: 66.70% versus 66.12% and 65.10%. There is no sufficiently broad evidence that their combination consistently helps across the other benchmark or acquisition conditions. Likewise, DQFT substantially outperforms component-wise DCT in the reported ablation. That is encouraging, but the paper needs stronger controls to establish whether the improvement comes from quaternion coupling, frequency selection, magnitude normalization, or differences between the respective aggregation implementations.
  3. The local orientation is derived from intersections between a circle and the contour, using nearby clockwise and counterclockwise intersection chords. The manuscript does not sufficiently establish how this construction behaves for highly concave leaves, irregular contours, multiple candidate intersections, or noisy segmentation boundaries. The implementation also needs to specify contour extraction, interpolation, handling of degenerate cases, and the exact patch-alignment procedure. These details matter because the boundary determines every subsequent feature patch.
  4. Appendix A.1 proves invariance to circular shifts of an already constructed descriptor sequence. This is a valid and useful property. It does not establish invariance to contour reversal, changes in contour sampling, local geometric perturbations, segmentation errors, or orientation-estimation noise. The manuscript should clearly separate these properties rather than allowing a mathematical invariance result to support a broader robustness interpretation.
Questions for authors
  1. What exact mathematical operation makes Quat.-4ary different from Real-4ary, beyond the use of quaternion notation?
  2. Are the real and quaternion variants matched in feature dimension, normalization, parameter count, and retained frequency information?
  3. Does the quaternion advantage persist when the transform axis is changed and when component-wise spectral processing is carefully controlled?
  4. How does the method handle contours with multiple intersections, highly concave shapes, and segmentation artifacts?
  5. Does the method remain effective when the contour direction is reversed, or is invariance intentionally limited to starting-point changes?
Reviewer_846ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the challenge of preserving spatial and structural information in training-free fine-grained leaf image retrieval by proposing Quaternion Structural-Spectral Aggregation (QSSA). The method extracts multi-scale boundary-aligned patches and employs dual-branch quaternion encoding to capture local spatial layouts and symmetry relationships from frozen CNN features. It then applies a discrete quaternion Fourier transform (DQFT) to aggregate contour-ordered descriptors, providing starting-point invariance and reducing sensitivity to boundary perturbations. Experiments on two leaf retrieval benchmarks, along with ablation studies, demonstrate the effectiveness of the proposed method.

Critical strengths
  1. Well-motivated problem. Training-free fine-grained leaf image retrieval is a practical and potentially valuable research problem.
  2. Well-structured and interpretable methodology. The proposed framework is clearly organized and integrates complementary structural representations.
Critical weaknesses
  1. Insufficient motivation. The paper does not sufficiently explain the importance and practical applications of fine-grained leaf image retrieval, making it difficult to fully assess the significance of the research problem.
  2. Lack of contour extraction details. The paper does not clearly describe how leaf contours are extracted.
  3. Outdated visual backbones. Most evaluated backbones are relatively outdated. The authors could consider more recent architectures, such as ConvNeXt, ConvNeXt V2, and DINOv3 (ConvNeXt variant). ViT-based backbones could also be explored by reshaping their spatial token features into 2D feature maps compatible with the proposed method.
  4. Insufficient cross-backbone aggregation comparisons. The experiments only evaluate QSSA across different backbones and compare aggregation methods using a fixed VGG-16 backbone. Without evaluating different aggregation methods across multiple backbones, it remains unclear whether QSSA consistently outperforms conventional aggregation methods and whether its improvements depend on specific backbone architectures.
  5. Concern about the contour starting-point invariance claim. The theoretical proof assumes that changing the contour starting point results in an exact circular shift of the descriptor sequence. However, I think this assumption may not always hold in practice, particularly when the starting point shifts by a non-integer multiple of the sampling interval, potentially producing different sampling locations and descriptors. Could such cases affect the effectiveness of the proposed method?
Questions for authors
  1. How are the leaf contours extracted? Could the authors provide more implementation details?
  2. Could the authors clarify whether the claimed contour starting-point invariance extends beyond exact circular shifts, particularly when the starting point is shifted by a non-integer multiple of the sampling interval?
Reviewer_49ywOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Quaternion Structural-Spectral Aggregation (QSSA) for training-free fine-grained leaf image retrieval. The authors argue that conventional aggregation of frozen CNN features often weakens the two-dimensional spatial organization within feature maps and does not jointly model how local structures evolve in sequence along the leaf contour. To address these limitations, QSSA first samples multi-scale, boundary-aligned patches along the leaf contour and then extracts their local features using a frozen CNN backbone. Within each feature channel, the method combines pooled responses from four quadrants into a quadrant-layout quaternion and constructs a second, symmetry-discrepancy quaternion from horizontal and vertical paired-region differences. These branches describe absolute regional layouts and pairwise regional relationships, respectively. The resulting quaternion descriptors are arranged in contour order and aggregated by a left-sided discrete quaternion Fourier transform (DQFT). Quaternion magnitudes remove variations caused by cyclic shifts in the arbitrary contour starting point, while low-frequency retention aims to suppress local boundary perturbations. Experiments are conducted on the four SoyCultivar200 subsets (Soy-Up, Soy-Mid, Soy-Low, and Soy-Joint) and PeanCultivar120, using the Bulls-eye score as the primary metric. QSSA achieves 68.23%, 66.26%, and 61.15% on Soy-Up, Soy-Mid, and Soy-Low, respectively, 92.52% on Soy-Joint, and 66.70% on PeanCultivar120. The paper also reports sensitivity analyses for sampling and frequency-retention parameters, comparisons of frozen backbones, component ablations, contour-perturbation experiments, and efficiency analyses.

Critical strengths

A training-free descriptor that jointly models local spatial layout and contour order. QSSA does not require additional optimization on the target dataset. It combines contour-aligned multi-scale patches, frozen CNN features, dual-branch quaternion encoding, and DQFT aggregation to preserve both spatial distributions within individual feature channels and ordered changes along the closed contour. This motivation is consistent with the subtle morphological differences and high visual similarity encountered in cultivar-level leaf retrieval. A relatively complete mathematical and structural formulation. The authors explicitly define the quadrant-based quaternion representation, paired-region symmetry discrepancies, and the left-sided DQFT convention. The appendix further derives contour starting-point invariance using the shift theorem and the multiplicative property of quaternion norms. The theoretical claim is limited to cyclic shifts of the contour starting point, rather than arbitrary deformations or reversals of contour direction; this scope is clarified in the appendix. Experimental evidence supporting the effectiveness of the method. QSSA outperforms SDePR on all three single-leaf subsets of SoyCultivar200 by 2.64, 1.98, and 1.24 percentage points, respectively. On PeanCultivar120, its score increases from SDePR’s 59.93% to 66.70%. Table 5 and Table A2 compare Real-4ary and Quat.-4ary encodings, the individual quaternion branches, and multiple aggregation alternatives. Appendix A.4 also examines noise added to contour coordinates. Thus, the paper does include ablation experiments, although the existing comparisons could better isolate certain mechanistic contributions.

Critical weaknesses

The independent advantage of quaternion-domain modeling over conventional spectral representations is not yet fully isolated. The comparison between Real-4ary (63.11%) and Quat.-4ary (66.70%) suggests that jointly processing quaternion components can be beneficial. Comparisons among Q-Avg (62.53%), Q-Max (62.19%), DCT (47.56%), and DQFT (66.70%) also support the role of spectral aggregation. However, the Real-4ary baseline applies separate real-valued DFTs to individual components, leaving open whether the gains arise more generally from coupled encoding and transformation. In particular, the paper does not directly compare quaternion operations with suitably coupled real-valued or complex-valued spectral approaches under matched descriptor dimensionality, normalization, and computational budgets. The existing evidence therefore supports the overall structural-spectral design more strongly than it isolates benefits unique to quaternion algebra. The statistical reliability of the main performance improvements needs stronger evidence. In Table 1, QSSA improves over SDePR on the three single-leaf subsets by 2.64, 1.98, and 1.24 percentage points, respectively, but is slightly worse on Soy-Joint (92.52% versus 92.64%). The manuscript reports point estimates without query-level confidence intervals or paired significance analyses for these comparisons. Given the modest gains on some subsets, query-level bootstrap confidence intervals or suitable paired tests for retrieval metrics would help assess the stability of the ranking improvements. This is not a request to repeat model training, since QSSA is training-free; the concern is uncertainty arising from the finite set of evaluation queries. The contour-error evaluation does not fully represent practical deployment conditions. Appendix A.4 perturbs leaf contour coordinates with Gaussian noise at 40, 35, and 30 dB. The Bulls-eye score decreases from 66.70% on clean contours to 65.20% at 30 dB, indicating some tolerance to this type of perturbation. However, simulated coordinate noise is not equivalent to realistic segmentation failures, such as missed contour segments, local adhesion, occlusions, or incomplete boundaries caused by cluttered backgrounds. The practical contour-extraction procedure and its failure modes are not described in sufficient detail to assess retrieval reliability when image preprocessing is unstable. Quantitative evaluation is also limited to two cultivar-level leaf benchmarks, leaving generalization to more diverse acquisition conditions less established. The parameter-selection rationale and accuracy–cost trade-off should be more transparent. Section 4.2 states that the default values N = 64, Ks = 4, and Kf = 4 are fixed a priori without target-dataset hyperparameter selection. Yet Table 3 shows that Kf = 8 achieves 67.97% on PeanCultivar120, compared with 66.70% for the default Kf = 4. Although the authors cite a reduction in descriptor dimensionality as a motivation, the manuscript does not sufficiently explain how the defaults were originally selected, whether they were based on prior knowledge, or how parameter-sensitivity experiments were kept separate from the training-free selection protocol. Appendix A.4 further reports 256 boundary patches per image, approximately 0.31 seconds for descriptor extraction, and 32 KiB of storage per image under the default configuration. A direct efficiency comparison with strong baselines under comparable hardware and implementation settings is still missing, making it difficult to assess the additional cost associated with the accuracy gains. These concerns do not establish data leakage; rather, they call for a clearer account of the evaluation protocol.

Questions for authors

Clarifications or additional evidence on the following questions could affect my assessment of the technical contribution and experimental reliability: Could the authors provide a more controlled comparison of spectral representations? With boundary sampling, frozen features, four-region inputs, descriptor length, distance metrics, and fusion procedures held constant, could QSSA be compared directly with complex-valued Fourier representations or coupled real-valued spectral representations to better quantify the contribution of quaternion algebra itself? Could the authors report the statistical uncertainty of the retrieval improvements? For the SoyCultivar200 subsets and PeanCultivar120, could the authors provide query-level confidence intervals or paired analyses for the main Bulls-eye score differences? Are the relatively small improvements on subsets such as Soy-Low stable, and what explains the slight underperformance relative to SDePR on Soy-Joint? How are contours extracted in practice, and does the method remain reliable under realistic errors? Are contours extracted automatically, and is manual correction required? How are occlusion, cluttered backgrounds, overlapping leaves, and gaps in segmentation handled? Could the authors evaluate retrieval with realistic segmentation outputs to complement the current simulated coordinate perturbations? How were the default parameters chosen, and how should the additional cost be assessed? How were N, Ks, and Kf fixed before evaluation without using target test data? Building on the runtime and storage measurements already reported in Appendix A.4, could the authors directly compare descriptor extraction time, storage, and query costs with strong baselines such as SDePR under comparable hardware conditions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.