acceptodds
Under review as a conference paper at ICLR 2027

HiGAS: Weakly-Supervised Oracle Bone Inscription Component Segmentation via Hierarchical Graph Alignment

Abstract

Oracle Bone Inscription (OBI) recognition is of great scientific importance for deciphering ancient Chinese characters, a large portion of which still remain undeciphered. Since complex OBI characters are usually composed of components with independent semantics, accurately uncovering the topological relationships between characters and components provides crucial support for character decipherment. However, existing methods suffer from high annotation costs, while significant structural variants of the same component make it challenging to match against standard forms. To address these issues, we propose HiGAS, a weakly supervised method for OBI component segmentation via hierarchical graph alignment. Using standard components as topological references, HiGAS constructs bottom-up hierarchical undirected graphs across segment, sub-stroke, and stroke levels for both input characters and reference components. Through within-level edge matching and cross-level topological priors, HiGAS effectively alleviates topological discontinuity and under-segmentation caused by local feature overfitting under weak supervision. Combined with a dynamic edge selection strategy and joint optimization of classification and segmentation completeness losses, the model predicts component presence probabilities while performing fine-grained edge-level segmentation. Extensive experiments on the public Component- and our extended Component- datasets demonstrate that HiGAS achieves a top- hit rate of in multi-component recognition while delivering segmentation results closest to expert annotations.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Weakly Supervised Learning, Graph Matching, Structural Representation Learning, Component Segmentation, Topological Alignment, Compositional Recognition, Oracle Bone Inscriptions

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_67sCOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to recognize and segment Oracle Bone Inscription (OBI) characters into semantic components. While the goal is to achieve component segmentation, the only available label is whether this component exists in the character. Due to the high annotation cost of pixel-level supervision, this paper proposes a weakly supervised method, HiGAS. HiGAS constructs bottom-up hierarchical undirected graphs at the segment, sub-stroke, and stroke levels for both input characters and components. It first establishes structural correspondences through within-level edge matching, and then propagates higher-level topological priors to enhance low-confidence segment-level responses. Together with structural completeness and hierarchical consistency objectives, the method aims to alleviate topological discontinuity and under-segmentation caused by local feature overfitting under image-level weak supervision.

Critical strengths
  1. The motivation for weakly-supervised setting is real.
  2. The results in Table 1 and Table 2 are very impressive. The improvement is large.
  3. Using loss in a hierarchical way seems to provide much better results.
Critical weaknesses
  1. There are a lot of manually designed features that lack ablation. For example, for higher-level strokes, there are covered segment count, cumulative and maximum inflection angles, boundary-closure proximity. I am wondering what if we use the naive GNN? Or inversely only use handcrafted features. It is unclear whether the performance mainly comes from the hierarchical reasoning or handcrafted priors.
  2. The classification loss applies the same image-level presence supervision independently at the segment, sub-stroke, stroke, and fused levels. However, this paper only ablates that this set of loss is necessary without ablating these hierarchical auxiliary losses individually, making it unclear whether supervision at every graph level is necessary, especially since higher-level scores already influence the final prediction through cross-level enhancement.
Questions for authors

In Eq.1, why use attention-like similarity instead of standard cosine similarity?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 67sC (Part 1/2: Feature Ablations & Similarity Design)

Dear Reviewer 67sC,

We sincerely thank you for your positive evaluation of our motivation, hierarchical loss design, and experimental results. We divide our response into two parts: Part 1/2 addresses the ablation on handcrafted attributes and GINE message passing (W1) and the similarity design in Equation (1) (Q1); Part 2/2 addresses the individual ablation of hierarchical auxiliary losses (W2). All corresponding updates are highlighted in blue in the revised PDF.

[W1] Ablation on Handcrafted Attributes and GINE Message Passing

There are a lot of manually designed features that lack ablation... What if we use the naive GNN? Or inversely only use handcrafted features. It is unclear whether the performance mainly comes from the hierarchical reasoning or handcrafted priors.

To separately evaluate the contributions of the handcrafted attributes and GINE-based neighborhood message passing, we conducted two additional ablation experiments while keeping the three-level graph structure and cross-level enhancement unchanged:

  1. Without handcrafted attributes (Naive GNN): Setting all node feature vectors to one and all edge attributes to zero at each graph level, retaining only graph connectivity and GINE message passing.
  2. Without GINE message passing (Handcrafted Only): Retaining all designed node and edge attributes across all three levels while replacing GINE neighbor aggregation with a per-node network without message passing.
SettingHit@1 (%)Hit@3 (%)Hit@5 (%)IoU (%)
Without handcrafted attributes37.0760.9670.6637.43
Without GINE message passing30.7858.5270.7438.03
HiGAS80.9794.0796.7764.35
  • Role of Handcrafted Attributes: Removing the designed node and edge attributes reduces Hit@1 from 80.97% to 37.07% and segmentation IoU from 64.35% to 37.43%. Without explicit geometric and topological attributes across levels, raw graph connectivity alone is insufficient to distinguish fine-grained stroke variations, showing that the designed attributes provide essential structural descriptions.
  • Role of GINE Message Passing: Retaining the handcrafted attributes while removing GINE message passing causes an even larger drop in Hit@1 (to 30.78%) and reduces IoU to 38.03%. Without GINE message passing to aggregate neighborhood context across connected nodes and edges, isolated handcrafted attributes cannot effectively resolve local structural ambiguities. Together, these results confirm that both the multi-level handcrafted attributes and GINE-based neighborhood message passing are indispensable and work jointly to achieve accurate recognition and segmentation.

(Updated in App. F, Table 4)

[Q1] Scaled Dot-Product Similarity vs. Cosine Similarity in Equation (1)

In Eq.1, why use attention-like similarity instead of standard cosine similarity?

We adopt a scaled dot product after learnable projections in Equation (1) primarily for two considerations regarding node matching and subsequent edge correspondence aggregation:

  1. Node Matching Flexibility: Equation (1) projects the node representations from both the character graph and the component graph into a shared matching space before computing their scaled dot product. Unlike standard cosine similarity (which normalizes vectors to unit length), the scaled dot product reflects both the direction and the magnitude of the projected representations, allowing the model to flexibly learn the matching strength of node pairs during training.
  2. Edge Endpoint Correspondence Aggregation: The node matching scores from Equation (1) directly contribute to the edge correspondence score in Equation (2), where the two possible endpoint alignment configurations of an undirected edge pair are summed and aggregated via the LogSumExp function. In LogSumExp aggregation, the score difference between the two endpoint configurations determines their relative contributions, where a larger difference allows the higher-scoring correspondence to dominate. While cosine similarity restricts node scores to [-1, 1], the unbounded scaled dot product provides greater flexibility in learning node score magnitudes and adjusting the relative contributions of competing endpoint correspondences.

(Updated in Sec. 3.3.1, Page 5)

AuthorsOpenReview ↗

Response to Reviewer 67sC (Part 2/2: Hierarchical Auxiliary Loss Ablation)

Dear Reviewer 67sC,

Below is Part 2/2 of our response, addressing your question regarding the individual contributions of hierarchical auxiliary classification losses (W2).

[W2] Individual Ablation of Hierarchical Auxiliary Losses in Lcls\mathcal{L}_{\mathrm{cls}}

The classification loss applies the same image-level presence supervision independently at the segment, sub-stroke, stroke, and fused levels. However, this paper only ablates that this set of loss is necessary without ablating these hierarchical auxiliary losses individually...

We agree that an overall ablation alone is insufficient to clarify the individual contribution of supervision at each graph level. In Equation (7), the classification loss Lcls\mathcal{L}_{\mathrm{cls}} consists of the fused-level classification loss Lfuse=BCE(σ(Sa,b),ya,b)\mathcal{L}_{\mathrm{fuse}} = \mathrm{BCE}(\sigma(S_{a,b}), y_{a,b}) and three level-specific auxiliary classification losses Lr=BCE(σ(Sa,br),ya,b)\mathcal{L}_{r} = \mathrm{BCE}(\sigma(S_{a,b}^{r}), y_{a,b}) for r∈seg,sub,strr \in \\{\mathrm{seg}, \mathrm{sub}, \mathrm{str}\\}:

Lcls=λmilLfuse+λsegLseg+λh2(Lsub+Lstr)\mathcal{L}_{\mathrm{cls}}=\lambda_{\mathrm{mil}}\mathcal{L}_{\mathrm{fuse}}+\lambda_{\mathrm{seg}}\mathcal{L}_{\mathrm{seg}}+\frac{\lambda_{h}}{2}(\mathcal{L}_{\mathrm{sub}}+\mathcal{L}_{\mathrm{str}})

While keeping the three-level graph structure, cross-level enhancement, and Lfuse\mathcal{L}_{\mathrm{fuse}} unchanged, we evaluated eight supervision settings: (1) independently removing each of the three auxiliary losses, (2) retaining Lfuse\mathcal{L}_{\mathrm{fuse}} with only one auxiliary loss at a time, and (3) retaining only Lfuse\mathcal{L}_{\mathrm{fuse}} without any auxiliary losses:

SettingHit@1 (%)Hit@3 (%)Hit@5 (%)IoU (%)Prec. (%)Rec. (%)F1 (%)
HiGAS80.9794.0796.7764.3592.6568.9277.58
Without Lseg\mathcal{L}_{\mathrm{seg}}80.0793.8996.6359.6487.6964.6172.70
Without Lsub\mathcal{L}_{\mathrm{sub}}79.4493.5596.5558.4284.1164.4471.16
Without Lstr\mathcal{L}_{\mathrm{str}}80.8394.0396.5359.6086.7665.4472.87
Fused + segment79.3693.0896.2458.6284.9464.8471.83
Fused + sub-stroke80.3593.4596.3159.6387.6764.5372.69
Fused + stroke79.6293.4996.4758.9386.1364.7171.78
Fused only78.6193.5096.3655.9484.5660.9568.96

Removing any single auxiliary loss reduces segmentation IoU by 4.71 to 5.93 percentage points, while retaining Lfuse\mathcal{L}_{\mathrm{fuse}} with only one auxiliary loss improves over using Lfuse\mathcal{L}_{\mathrm{fuse}} alone (55.94% IoU) but remains below the full model (64.35% IoU). This demonstrates that although cross-level enhancement propagates higher-level scores downward, such interaction cannot fully replace direct classification supervision at individual graph levels. Applying auxiliary losses across all three levels provides complementary supervision that enables the model to better exploit multi-level structural information.

(Updated in App. F, Table 5)

Thanks for the detailed rebuttal. My concerns have been addressed.

AuthorsOpenReview ↗

Thank You for Your Positive Feedback

Dear Reviewer 67sC,

Thank you for confirming that our rebuttal has addressed your concerns. As the discussion deadline approaches, if you feel that our clarifications and additional ablations strengthen the paper, we would be grateful if you would consider updating your rating or increasing your confidence level in the final evaluation.

Reviewer_9CD7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes HiGAS, a weakly supervised method for OBI component segmentation via hierarchical graph alignment. Using standard components as topological references, HiGAS constructs bottom-up hierarchical undirected graphs across segment, sub-stroke, and stroke levels for both input characters and reference components. Through within-level edge matching and cross-level topological priors, HiGAS effectively alleviates topological discontinuity and under-segmentation caused by local feature overfitting under weak supervision. Combined with a dynamic edge selection strategy and joint optimization of classification and segmentation completeness losses, the model predicts component presence probabilities while performing fine-grained edge-level segmentation.

Critical strengths
  1. The hierarchical alignment design is well suited to matching structurally variable components. The ablations support the contribution of cross-level enhancement and completeness losses to recovering component strokes under weak supervision.

  2. The architecture diagram and qualitative comparisons help explain the method. The paper describes the graph construction and matching procedure clearly enough to follow the main design.

Critical weaknesses
  1. The evaluation does not measure complete multi-component segmentation. Component-50 evaluates component presence, while Component-20 provides masks for only one target component per image. Segmentation scores are also conditional on passing the presence threshold, so they do not capture errors across the full prediction and segmentation pipeline.

  2. The baseline configurations are unclear. CL-OBIS uses both pixel annotations and weakly annotated data with stroke constraints [1], while P³ includes a mask-generation and filtering pipeline [2]. This paper does not explain which procedures and training resources were retained, making the segmentation comparison difficult to assess.

  3. The Component-50 split does not specify whether character categories overlap between training and testing. Without this information, the results cannot distinguish generalization to new character categories from recognition of additional examples of familiar categories.

  4. The interpretation of Hit@5 overstates component coverage. This metric requires only one correct component among five candidates. The reported complete recall of 62.18% for four-component characters therefore does not support the statement that almost all true components are retrieved within five candidates.

[1] Zhikai Hu et al. Component-Level Segmentation for Oracle Bone Inscription Decipherment. AAAI, 2025.

[2] Pengjie Wang et al. Puzzle Pieces Picker: Deciphering Ancient Chinese Characters with Radical Reconstruction. ICDAR, 2024.

Questions for authors
  1. What is the segmentation performance on all 304 test images when cases rejected by the presence predictor are included? Are any quantitative results available for jointly identifying and segmenting all components in a character?

  2. What training data and adaptations were used for CL-OBIS and P³? In particular, were CL-OBIS’s weakly supervised training and stroke constraints retained, and how were P³’s component masks obtained?

  3. How was the Component-50 split constructed, and how many character categories are shared between training and testing? If the test set contains unseen categories, can their results be reported separately?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 9CD7 (Part 1/2: Full-Set & Multi-Component Segmentation)

Dear Reviewer 9CD7,

We sincerely thank you for your constructive review and insightful questions. To ensure clarity and readability, we organize our response into two parts (Part 1/2 addresses W1 / Q1, and Part 2/2 addresses W2–W4 / Q2–Q3). All corresponding updates are highlighted in blue in the revised PDF.


[W1 & Q1] End-to-End Evaluation on All 304 Images and Multi-Component Segmentation

What is the segmentation performance on all 304 test images when cases rejected by the presence predictor are included? Are any quantitative results available for jointly identifying and segmenting all components in a character?

1. End-to-End Pipeline Evaluation on All 304 Test Images:

In the inference pipeline of HiGAS, the model first predicts the component presence probability (y^a,b\hat{y}_{a,b}) and subsequently performs fine-grained segmentation on components predicted as present (y^a,b>δcls\hat{y}_{a,b} > \delta_{\mathrm{cls}}). In Section 4.6 of the original manuscript, we evaluated segmentation on the predicted-present subset (283 targets with y^a,b>δcls\hat{y}_{a,b} > \delta_{\mathrm{cls}}, covering 93.0993.09\\% of the test set) to assess segmentation mask quality independently of presence prediction errors. To evaluate the complete end-to-end pipeline across all 304 test targets in Component-20, we now include the 21 targets rejected by the presence predictor (y^a,b≤δcls\hat{y}_{a,b} \leq \delta_{\mathrm{cls}}) by treating them as absent and setting their predicted masks to empty masks (00 score across all metrics), denoted as HiGAS (end-to-end) in Table 1(b):

Evaluation ProtocolScopeIoU (%)Prec. (%)Rec. (%)F1 (%)
Predicted-present subset (y^a,b>δcls\hat{y}_{a,b} > \delta_{\mathrm{cls}})283 detected targets64.3592.6568.9277.58
HiGAS (end-to-end)All 304 test targets59.9186.1164.1672.22

Incorporating false negatives from the presence prediction stage naturally lowers the overall scores, yet HiGAS still achieves 59.91% IoU, 86.11% Precision, and 72.22% F1 across all 304 test images, remaining substantially higher than the segmentation baselines reported in Table 1(b) (34.81% IoU for CL-OBIS and 42.25% IoU for P3P^3).

2. Multi-Component Joint Recognition and Segmentation:

Because the original Component-20 test set (304 images) annotated only one target component per image across 20 categories, we expanded its ground truth by manually adding pixel-level masks and category labels for all components from our 50 target categories. This increased the total annotated components from 304 to 494 components (+190 newly annotated components), with 163 test images containing ≥2\geq 2 annotated components. We evaluate HiGAS across three scopes (with combined character-level masks evaluated in App. G, Table 7):

  1. Single-component scope (494 components): Evaluating all 494 individual annotated components across the full test set to measure component-level detection rate and segmentation accuracy.
  2. Whole-character scope (304 images): Evaluating all 304 test images at the character level, where an image is counted as detected only when all of its annotated components are simultaneously detected.
  3. Multi-component subset (163 images): Evaluating complete character-level detection and joint recognition and segmentation specifically on images containing ≥2\geq 2 annotated components.
Evaluation ScopeUnitsDet. Rate (%)IoU (%)Prec. (%)Rec. (%)F1 (%)
All cases; component-level detection494 components92.91 (459/494)67.0992.2072.1179.73
All components detected; full set304 images89.14 (271/304)68.1392.3173.2080.56
All components detected; multi-component set163 images85.89 (140/163)70.4692.2475.5382.20

HiGAS detects 92.91% of all individual components (67.09% IoU), recovers all constituent components simultaneously in 89.14% of all 304 test images (68.13% IoU), and achieves an 85.89% complete detection rate with 70.46% IoU and 82.20% F1 (Precision >92>92\\%) on the 163 multi-component characters. This confirms that HiGAS reliably performs joint recognition and segmentation of multiple coexisting components within complex characters.

(Updated in Sec. 4.6, Table 1(b), Pages 8–9 & App. G, Table 7)

AuthorsOpenReview ↗

Response to Reviewer 9CD7 (Part 2/2: Baselines, Data Split & Metrics)

Dear Reviewer 9CD7,

Below is Part 2/2 of our response, addressing your questions regarding baseline configurations (W2 / Q2), Component-50 data splitting (W3 / Q3), and the distinction between Hit@5 and CR@5 (W4).


[W2 & Q2] Baseline Configurations for CL-OBIS and P3P^3

What training data and adaptations were used for CL-OBIS and P3P^3? In particular, were CL-OBIS's weakly supervised training and stroke constraints retained, and how were P3P^3's component masks obtained?

1. Training Configurations of CL-OBIS: We retained its original weak supervision and stroke constraints and evaluated two settings:

  • CL-OBIS-C20 (reported in paper): 9,245 weakly labeled and 708 pixel-annotated Component-20 training images.
  • CL-OBIS-C50 (expanded weak data): 27,394 Component-50 weakly labeled images containing at least one target component, plus the same 708 pixel-annotated images.
Baseline SettingIoU (%)Prec. (%)Rec. (%)F1 (%)
CL-OBIS-C2034.8172.0441.0850.93
CL-OBIS-C5052.0054.3690.3465.06
Full-character foreground48.5648.56100.0062.28

CL-OBIS-C50's high Recall (90.34%) but low Precision (54.36%) indicates over-segmentation, likely due to the imbalance between 27,394 weakly labeled and 708 pixel-annotated images. Its results resemble the full-character foreground reference (48.56% Precision, 100% Recall), suggesting whole-character activation rather than component isolation. We therefore report CL-OBIS-C20.

2. Implementation Details of P3P^3: We generated candidate masks using three parallel branches: OpenCV contour extraction, a pretrained SAM fine-tuned on the 708 training images, and U-Net trained on the same images. A pretrained MoCo encoder extracted candidate features, which KNN matched against a class-specific reference bank of standard components from Yinqi Wenyuan. The highest-similarity candidate was selected as the segmentation result.

(Updated in Sec. 4.3, Page 8 & App. E, Table 3)


[W3 & Q3] Component-50 Split and Unseen Character Generalization

How was the Component-50 split constructed, and how many character categories are shared between training and testing? If the test set contains unseen categories, can their results be reported separately?

Component-50 (78,819 images across 1,188 character categories) was randomly split at the image level into 66,996 training and 11,823 test images after two-stage deduplication. Examination of the split shows that 1,165 character categories (10,728 test images) are shared between training and testing, whereas 23 character categories (1,095 test images) are completely absent from the training set:

SubsetCategoriesImagesHit@1 (%)Hit@3 (%)Hit@5 (%)
Full test1,18811,82380.9794.0796.77
Seen categories1,16510,72881.0994.2496.84
Unseen categories231,09579.8292.4296.07

HiGAS maintains comparable performance on unseen categories, with only 0.70–1.65 percentage-point drops across Hit@1/3/5 compared with the full test set, demonstrating strong generalization to novel character structures.

(Updated in Secs. 4.3 and 4.5, Table 1(a), Page 8; App. G, Table 6)


[W4] Distinction Between Hit@5 and Complete Recall (CR@5)

The interpretation of Hit@5 overstates component coverage. This metric requires only one correct component among five candidates.

We agree that Hit@5 (requiring at least one true component in the top-5) cannot support the claim that all ground-truth components are retrieved. This requires CR@5, which measures the proportion of characters for which all ground-truth components are included in the top-5 candidates.

As defined in Sec. 4.2 and Table 2, complete recall is denoted as CR@(n+m)\mathrm{CR@}(n+m), where nn is the number of ground-truth components and mm is the additional candidate budget. Thus, top-5 complete recall corresponds to n+m=5n+m=5. HiGAS achieves 93.87%, 88.97%, 81.05%, and 62.18% for characters with 1, 2, 3, and 4 components, respectively. Across all 11,823 test images, all ground-truth components are recovered within the top five in 10,671 images (90.26% CR@5). For the four-component subset, this is 74/119 images (62.18%), representing only 1.01% of the test set.

We clarified this distinction in Sec. 4.5 (Pages 8–9): Hit@5 indicates that at least one true component appears in the top five (96.77%), whereas CR@5 measures complete recovery of all true components (90.26%).

(Updated in Sec. 4.5, Pages 8–9)

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer 9CD7,

As the discussion deadline approaches, we would like to follow up and ask whether our responses and additional experiments on full-set end-to-end segmentation, multi-component joint segmentation, baseline configurations, unseen category generalization, and the CR@55 clarification have addressed your concerns.

Please let us know if any further clarification is needed. If your concerns have been resolved, we would be grateful if you would consider updating your rating.

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

This paper proposes HiGAS, a weakly supervised method for Oracle Bone Inscription (OBI) component recognition and segmentation trained using only image-level component presence labels. The model represents both characters and standard reference components using three hierarchical graphs: the Segment Graph, Sub-stroke Graph, and Stroke Graph. It encodes these graphs using GINE and performs within-level edge matching. Through cross-level structural enhancement, dynamic edge selection, and a joint objective combining component presence classification and segmentation completeness, HiGAS predicts component presence and produces fine-grained segmentation results.

Critical strengths
  1. The paper addresses a practically relevant problem in weakly supervised OBI component analysis, jointly performing component recognition and fine-grained segmentation using only image-level component presence labels for training. This setting reduces reliance on pixel- or box-level annotations and offers a valuable direction for reducing annotation costs in structural analysis of OBIs.
  2. The hierarchical graph representation is well motivated. The Segment Graph, Sub-stroke Graph, and Stroke Graph capture local geometric details, stroke continuity, and more complete stroke structures, respectively. Together with standard reference components, cross-level structural enhancement transfers higher-level structural matching evidence to segment edges, providing a reasonable mechanism for mitigating incomplete segmentation caused by reliance on locally discriminative structures under weak supervision.
  3. The reported results show strong performance on both component presence prediction and segmentation. On Component-50, HiGAS achieves a Hit@1 of 80.97%, compared with 67.68% for the DINOv2-based baseline. On the high-confidence subset used for segmentation evaluation on Component-20, it achieves an IoU of 64.35% and an F1-Score of 77.58%, outperforming the compared segmentation baselines. These results provide initial empirical support for the method’s effectiveness.
  4. The ablation study helps clarify the contributions of the hierarchical structure and completeness constraints. The full model achieves an IoU of 64.35%, whereas using only the Segment Graph or only the classification loss reduces IoU to 51.25% and 44.47%, respectively. Moreover, removing the structural topology completeness loss slightly improves some recognition metrics but substantially degrades segmentation performance, illustrating that optimizing component presence prediction alone is insufficient to ensure segmentation completeness.
Critical weaknesses
  1. The paper does not clearly describe the annotation procedure for Component-50. For example, it remains unclear how character categories in Yinqi Wenyuan are mapped to images in EVOBC. The paper also does not specify who performed the annotation, whether automated programs or models were used, or whether the annotations underwent expert verification.
  2. The paper does not clearly describe the splitting and deduplication procedures. This makes it difficult to determine whether the results reflect recognition performance on seen character categories or generalization to unseen character categories. The authors do not explain how identical or near-duplicate images are excluded across the Component-50 training and test splits, or between the Component-50 training split and the Component-20 test split.
  3. Several important hyperparameter values and implementation settings are not specified, including γ\gamma in Equation (9), the threshold for determining whether edges are “approximately collinear,” and the number or proportion of negative samples.
  4. Real-world OBI scans may contain missing, fragmented, or blurred strokes, which can introduce noise into the graphs at different hierarchical levels. However, the paper does not evaluate the model’s robustness to such noise, leaving its applicability to real-world scenarios insufficiently established.
Questions for authors
  1. Could the authors provide a more detailed account of how Component-50 was constructed? Specifically, how were the character–component correspondences from Yinqi Wenyuan mapped to images in EVOBC? Please also clarify who performed the annotation, whether automated programs or models were used, and whether the annotations underwent expert verification.
  2. Regarding Weakness 2, could the authors describe the data splitting and deduplication procedures? Please clarify whether Component-50 was split at the image level or the character-category level, and explain how identical and near-duplicate images were detected. In addition, how did the authors verify that there is no image overlap between the Component-50 training split and the Component-20 test split?
  3. Regarding Weakness 3, could the authors report the values of all important hyperparameters and implementation settings, including those listed above, to improve reproducibility?
  4. Regarding Weakness 4, could the authors report recognition and segmentation performance under varying degrees of missing strokes, stroke fragmentation, and spurious connections? A comparison between the single-level and three-level models under these perturbations would help establish whether the hierarchical structure mitigates or amplifies noise introduced during graph construction.
4 comments
AuthorsOpenReview ↗

Response to Reviewer eqWr (Part 1/2: Dataset & Deduplication)

Dear Reviewer eqWr,

We sincerely thank you for your positive evaluation of our work and your constructive feedback. To ensure clarity and readability, we organize our response into two parts (Part 1/2 addresses W1–W2 / Q1–Q2, and Part 2/2 addresses W3–W4 / Q3–Q4). All corresponding updates are highlighted in blue in the revised PDF.


[W1 & Q1] Construction and Annotation of Component-50

How were character-component correspondences from Yinqi Wenyuan mapped to images in EVOBC? Please clarify who performed the annotation, whether automated programs or models were used, and whether the annotations underwent expert verification.

Component-50 was constructed via a two-stage procedure combining automated script-based label transfer and expert cross-verification, without using any predictive models:

  1. Automated Script-Based Label Transfer: Character entries in Yinqi Wenyuan provide component-level decompositions for the 50 target components, and both Yinqi Wenyuan and EVOBC annotate OBI characters with their corresponding modern Chinese characters. Using an automated matching script, we matched OBI character images in EVOBC to Yinqi Wenyuan entries via their shared modern Chinese character labels, automatically transferring the 50-category component-presence annotations to the matched EVOBC images.
  2. Expert Cross-Verification: Because historical glyph variants may alter or omit certain components, two OBI experts (each with 5–7 years of paleographic research experience) independently examined and cross-verified the image–label pairs. Only images with unanimous expert consensus on component composition were retained, while inconsistent variant-form images were removed. (Updated in Sec. 4.1, Page 7)

[W2 & Q2] Deduplication, Data Splitting, and Unseen Category Generalization

Please clarify whether Component-50 was split at the image or character-category level, how identical and near-duplicate images were detected, and how you verified that no image overlap exists between the Component-50 training split and the Component-20 test split.

1. Two-Stage Deduplication (Performed Prior to Data Splitting): During dataset preparation (documented in Sec. 4.1, Page 7 and illustrated with visual examples in App. B, Fig. 1), we executed a two-stage deduplication procedure:

  • Within Component-50: We first applied SHA-256 hashing over decoded white-background RGB pixels to identify and remove exact duplicate images, followed by perceptual hashing (pHash) (with normalized foreground mask IoU screening) to remove near-duplicate images containing glyphs with highly similar shapes.
  • Between Component-50 and Component-20: We applied the exact same two-stage deduplication (SHA-256 + pHash) between Component-50 and Component-20 to remove any exact or near-duplicate images of Component-20 from Component-50. This strictly ensures that no test images from the Component-20 segmentation benchmark appear in the Component-50 training set.

2. Splitting & Unseen Character Generalization: Component-50 was then randomly split at the image level into 66,996 training and 11,823 test images. Examination of the split revealed that 1,095 test images spanning 23 character categories are completely absent from the training set. We evaluated HiGAS specifically on this unseen subset (HiGAS (unseen)):

Method / SubsetTest ImagesHit@1 (%)Hit@3 (%)Hit@5 (%)
CL-OBIR (Overall)11,82354.2876.0183.46
DINOv2-based baseline (Overall)11,82367.6887.9692.92
HiGAS (unseen: 23 categories)1,09579.8292.4296.07
HiGAS (Ours, Overall)11,82380.9794.0796.77

Compared with its overall performance, HiGAS exhibits only a slight decline on unseen character categories (-1.15% in Hit@1 and -0.70% in Hit@5) while still outperforming both CL-OBIR and DINOv2 by substantial margins. This demonstrates that HiGAS learns generalizable component-level topological alignments rather than memorizing training character categories. (Updated in Sec. 4.1, Sec. 4.3, Sec. 4.5, Table 1(a), Pages 7–8 & App. B, Fig. 1, App. G, Table 6)

AuthorsOpenReview ↗

Response to Reviewer eqWr (Part 2/2: Hyperparameters & Robustness)

Dear Reviewer eqWr,

Below is Part 2/2 of our response, addressing reproducibility (W3/Q3) and robustness to topological perturbations (W4/Q4).

[W3 & Q3] Hyperparameters, Implementation Settings, and Code Release

Could the authors report the values of all important hyperparameters and implementation settings, including those in Equation (9), the collinearity threshold, and negative sampling, to improve reproducibility?

Parameters in Equation (9) (Lgrp\mathcal{L}_{\mathrm{grp}}): γ\gamma is set to 00, the numerical stability constant is ε=10−6\varepsilon = 10^{-6} to prevent division by zero, and the loss weight in Eq. (10) is λgrp=0.05\lambda_{\mathrm{grp}} = 0.05.

Approximate Collinearity Criterion (Sec. 3.2.2 & App. A): Following the stroke continuity analysis near junctions in [Xu et al., 2024], for any two candidate edges sharing a junction point, we sample a point 6 pixels away from the junction along each edge to construct outward unit direction vectors di\mathbf{d}_i and dj\mathbf{d}_j. We then compute the opposing alignment score a=−di⊤dj∈[−1,1]a = -\mathbf{d}_i^\top \mathbf{d}_j \in [-1, 1], where a→1a \to 1 indicates alignment along the same straight line in opposite directions. Two edges are merged as approximately collinear when a>0.8a > 0.8.

Negative Sampling Strategy (Sec. 4.4 & App. C): For each character image, we retain all annotated positive components and dynamically sample an equal number of negative component categories from the remaining absent categories at each training epoch, resulting in a 1:1 positive-to-negative ratio.

Complete Hyperparameter Summary & Code Release: We have compiled a comprehensive summary of all model, graph-extraction, and training hyperparameters in App. C (Table 1). We will publicly release our complete source code, pretrained weights, and dataset annotations upon publication to ensure full reproducibility.

(Updated in Sec. 3.2.2, Sec. 3.4.2, Sec. 4.4, Pages 4, 7, 8 & App. A, App. C(Table 1))

[W4 & Q4] Robustness to Missing Strokes, Fragmentation, and Spurious Connections

Could the authors report recognition and segmentation performance under varying degrees of missing strokes, stroke fragmentation, and spurious connections, comparing the single-level and three-level models under these perturbations?

We randomly perturbed Component-20 test images at 5%, 10%, and 15% using three equally likely perturbations before graph construction: stroke deletion (removing segments between junctions and endpoints), stroke breaks (removing short interior segments), and spurious connections (adding artificial connections). Examples are shown in App. D, Fig. 2. Both HiGAS and the single-level model (GsegG^{\mathrm{seg}} only) were evaluated on the same perturbed image sets. Pass Rate denotes the percentage of image–target pairs with predicted presence probability y^a,b>0.5\hat{y}_{a,b} > 0.5.

ModelPerturbation RatioPass Rate (%)IoU (%)Prec. (%)Rec. (%)F1 (%)
HiGAS (Three-level)0%93.0964.3592.6568.9277.58
HiGAS (Three-level)5%92.4364.1292.1168.8277.33
HiGAS (Three-level)10%92.1163.6992.0468.3976.98
HiGAS (Three-level)15%91.4563.5492.0668.2576.85
Only GsegG^{\mathrm{seg}} (Single-level)0%90.1351.2586.4154.7664.20
Only GsegG^{\mathrm{seg}} (Single-level)5%90.1351.0286.3654.5664.00
Only GsegG^{\mathrm{seg}} (Single-level)10%89.4750.8386.2754.4363.85
Only GsegG^{\mathrm{seg}} (Single-level)15%89.1450.7786.3554.3563.79

Key Observations: As perturbation increases from 0% to 15%, HiGAS's Pass Rate, IoU, and F1 decrease by only 1.64, 0.81, and 0.73 percentage points, respectively, compared with 0.99, 0.48, and 0.41 points for the single-level model. This indicates slightly greater sensitivity of hierarchical modeling, as local perturbations may affect inter-level merging during graph construction. Nevertheless, HiGAS consistently outperforms the single-level model at all perturbation levels. Even at 15% perturbation, it achieves 63.54% IoU and 76.85% F1, exceeding the unperturbed single-level model by 12.29 and 12.65 percentage points, respectively.

(Updated in Sec. 4.7, Table 3, Pages 9–10; App. D, Fig. 2 and Table 2)

This rebuttal has addressed my questions, so I will maintain my score and increase my confidence level.

AuthorsOpenReview ↗

Thank You for Your Feedback and Support

Dear Reviewer eqWr,

Thank you for reviewing our response and confirming that all of your questions have been addressed. We appreciate your constructive feedback, which has helped improve our manuscript, as well as your support in increasing your confidence level.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Area Chair and All Reviewers

Dear Area Chair and Reviewers,

We thank the reviewers for their constructive feedback and for recognizing the importance of weakly supervised OBI component analysis, our hierarchical graph design, and the improvements in recognition and segmentation. We have revised the manuscript, added experiments, and highlighted all changes in blue. The main updates are:

Dataset Construction, Deduplication, and Unseen Category Generalization (Reviewer eqWr W1–W2/Q1–Q2, Reviewer 9CD7 W3/Q3)

We clarified the Component-50 construction, including script-based label transfer via shared modern Chinese character labels and independent verification by two OBI experts (Sec. 4.1, Page 7). We also documented the two-stage deduplication (SHA-256 and perceptual hashing) performed during dataset preparation, ensuring no Component-20 test images enter Component-50 training (Sec. 4.1, Page 7; App. B, Fig. 1). On 1,095 test images from 23 unseen character categories, HiGAS achieves 79.82%/92.42%/96.07% Hit@1/3/5, demonstrating strong generalization to unseen character structures (Secs. 4.3, 4.5; Table 1(a), Page 8; App. G, Table 6).

Robustness Under Synthetic Stroke Perturbations (Reviewer eqWr W4/Q4)

We evaluated HiGAS and the single-level model (Only GsegG^{\mathrm{seg}}) under three types of synthetic stroke perturbations at 5%, 10%, and 15% (Sec. 4.7, Table 3, Pages 9–10; App. D, Fig. 2, Table 2). As perturbation increases from 0% to 15%, HiGAS's IoU and F1 decrease by only 0.81 and 0.73 percentage points, respectively. Even at 15% perturbation, HiGAS outperforms the unperturbed single-level model by 12.29 IoU and 12.65 F1 percentage points, demonstrating robust hierarchical modeling (App. D, Table 2).

Ablations on Handcrafted Attributes, GINE Message Passing, and Auxiliary Losses (Reviewer 67sC W1–W2/Q1)

Ablations show that removing handcrafted attributes or GINE message passing reduces Hit@1 to 37.07% or 30.78%, respectively, with IoU dropping to 37.43% or 38.03%, confirming their complementary contributions (App. F, Table 4). In an eight-setting auxiliary-loss ablation, removing any single auxiliary loss reduces IoU by 4.71–5.93 percentage points, while retaining only the fused-level loss reduces IoU from 64.35% to 55.94%, demonstrating the importance of supervision across graph levels (App. F, Table 5).

End-to-End Full-Set Evaluation and Multi-Component Joint Segmentation (Reviewer 9CD7 W1/Q1)

We evaluated the complete end-to-end pipeline on all 304 Component-20 test targets, assigning empty masks to targets rejected by the presence predictor (y^a,b≤δcls\hat{y}_{a,b} \leq \delta_{\mathrm{cls}}). HiGAS achieves 59.91% IoU, 86.11% Precision, and 72.22% F1, accounting for false negatives in the presence prediction stage (Sec. 4.6, Table 1(b), Pages 8–9). We also expanded pixel-level annotations to cover all 50 component categories, yielding 494 annotated components, including 163 multi-component characters. HiGAS detects 92.91% of individual components (67.09% IoU), recovers all components in 89.14% of test images (68.13% IoU), and achieves 85.89% complete detection on multi-component characters (70.46% IoU, 82.20% F1), demonstrating effective joint recognition and segmentation (Sec. 4.6, Page 9; App. G, Table 7).

Baseline Protocols, Implementation Details, and Metric Clarifications (Reviewer eqWr W3/Q3, Reviewer 9CD7 W2/W4/Q2, Reviewer 67sC Q1)

We detailed the training configurations of CL-OBIS and P3P^3 along with an analysis of CL-OBIS trained on expanded weak data from Component-50 (Sec. 4.3, Page 8 & App. E, Table 3), added all hyperparameter values and collinearity criteria (Sec. 3.2.2, Sec. 3.4.2, Sec. 4.4, Pages 4, 7, 8 & App. A, App. C, Table 1), explained the projected scaled dot product in Eq. (1) (Sec. 3.3.1, Page 5), and explicitly distinguished Hit@5 from Complete Recall (CR@5), where HiGAS achieves an overall CR@5 of 90.26% across all 11,823 test images (Sec. 4.5, Pages 8–9).

Please refer to our point-by-point responses below for detailed tables and discussions. We remain at your disposal for any further questions during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.