SASP: Strip-Aware Spatial Perception for Fine-Grained Bird Image Classification

Zheng Wang

REVIEW 4 major objections 6 minor 11 references

SASP: Strip-Aware Spatial Perception for Fine-Grained Bird Image Classification

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that strip-aware spatial perception (SASP) lifts fine-grained bird classification to 72.58% on CUB-200-2011, surpassing decade-old baselines by fusing row/column strip pooling with channel reweighting on a ResNet-50…

desk verdict Strip pooling plus channel attention, compared only against 2013–2015 baselines, does not support the claimed 72.58% as a substantial gain on CUB-200-2011. read the letter →

arxiv 2505.24380 v2 pith:BNA45OZC submitted 2025-05-30 cs.CV cs.AI

Zheng Wang This is my paper

classification cs.CVcs.AI

keywords fine-grainedbirdclassificationstrippoolinglong-rangespatialdependencieschannelattentionCUB-200-2011ResNet-50perceptionvisualrecognition

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes SASP, a bird classification model built on ResNet-50, and claims it reaches 72.58% accuracy on the CUB-200-2011 dataset. The central claim is that two new modules, an extensional perception aggregator (EPA) and a channel semantic weaving (CSW) layer, let the model capture long-range, strip-like structures such as outstretched wings and tails while suppressing background clutter. The paper reports that this result substantially outperforms older baselines in its comparison table, notably GoogLeNet at 68.19%. A sympathetic reader would take the paper as arguing that combining directional spatial pooling with channel attention is an effective, efficient recipe for fine-grained bird recognition.

What carries the argument

The key machinery is the extensional perception aggregator (EPA), a five-branch module that combines local detail with horizontal and vertical strip pooling: the horizontal branch pools each row, the vertical branch pools each column, and two asymmetric branches use 1x3 and 3x1 convolutions on pooled features, with outputs fused and added residually to the input feature map. This is followed by channel semantic weaving (CSW), a channel-attention layer that generates a global descriptor via average pooling, passes it through a two-layer bottleneck, and multiplies the feature map by sigmoid-derived per-channel weights. The combination lets the network attend to elongated structures and recalibrate feature channels without heavy computational overhead.

What would settle it

Run SASP on CUB-200-2011 without the automatic face-cropping step, and also run a current high-accuracy fine-grained classifier on the same train/test split with identical training settings; if SASP's accuracy drops well below modern results or changes materially with the crop preprocessing, the paper's central performance claim is refuted.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a ResNet-50 backbone, pruned before its final residual stage and augmented with strip-aware spatial perception, can achieve 72.58% classification accuracy on CUB-200-2011 with a lightweight head. The EPA module uses parallel branches of local 3x3 convolution, horizontal and vertical strip pooling, and two asymmetric strip convolutions to model both fine texture and long-range cross-row/cross-column dependencies. The CSW layer then applies channel-wise reweighting through a compact bottleneck and sigmoid activation to amplify discriminative channels and suppress noise. Together these components enable what the paper calls leap-wise perception of elongated bird structures across pose, scale, and background variation.

Load-bearing premise

The load-bearing premise is that comparing SASP to the 2013-2015 baselines in Table 2, under the face-cropping protocol described, is a fair and reproducible benchmark; if modern baselines or an unspecified face detector were used, the claimed substantial gain would not hold.

Editorial extensions

If this is right

If SASP's accuracy claim holds under fair comparison, strip pooling along rows and columns is a viable way to capture elongated bird parts that square pooling or fixed kernels miss.
The CSW channel reweighting provides a lightweight complement to spatial attention, suggesting that channel-wise calibration can boost fine-grained discrimination without adding many parameters.
The model's end-to-end design on a truncated ResNet-50 indicates that high accuracy can be achieved with relatively modest architectural modifications.
The reported convergence behavior, with rapid early loss decrease followed by stable refinement, supports the interpretation that EPA captures spatial structure first and CSW refines channel semantics during training.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Inference beyond the paper: the 72.58% figure would likely be much lower relative to modern fine-grained classifiers, which commonly exceed 90% on CUB-200-2011; thus the 'substantial gain' is really a claim against the 2013-2015 baselines listed in the table.
Inference beyond the paper: the face-cropping preprocessing described as 'automatically detected and cropped' is not fully specified, so the claimed gain could hinge on the choice of face detector; an ablation without cropping or with a publicly described detector would make the result reproducible.
Inference beyond the paper: the strip-pooling idea could transfer to other domains with elongated structures, such as text line detection, road segmentation, or medical vessel tracking, where row/column dependencies dominate.
Inference beyond the paper: the interpretability claim could be tested by verifying whether EPA's learned strip responses align with birds' wings and tails and whether CSW weights concentrate on plumage-relevant channels.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper is a combination of strip pooling and channel attention dressed up as new modules, and its only experimental comparison is against a 2013–2015 baseline set. The reported 72.58% on CUB-200-2011 is far below modern results, so the 'substantial performance gain' claim doesn't hold up. That's the one thing to know.

To be fair, the paper is clearly written and the architecture is described carefully. The idea of adding strip pooling branches to a ResNet-50 and then recalibrating channels is coherent, and if the comparison were fair—say, against a plain ResNet-50 trained under the same protocol—the modules might even show a modest gain. But the paper doesn't run that comparison. Table 2 includes GoogLeNet and older hand-crafted methods, all from 2013–2015, and no modern fine-grained method. The abstract and Section 3.3 claim significant improvement, but that's an artifact of the baseline selection.

The other substantive problems: no ablations, no error bars, no code, and the face-crop preprocessing in Section 3.2 says 'automatically detected' without naming the detector or whether part annotations were used. That makes the result irreproducible and potentially unfair. And the two modules are essentially the strip pooling of Hou et al. (ECCV 2020) and the channel attention of Hu et al. (CVPR 2018), neither of which is cited. The combination is not a novel contribution as presented.

I would not send this to peer review. It's an incremental recombination with an unsupported comparative claim. The authors need to redo the experiments against modern baselines, add ablations, cite prior work, and specify the preprocessing. If the modules still help under those conditions, there's potentially a workshop-level result. But as is, desk reject.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes SASP, a fine-grained bird image classification framework built on a ResNet-50 backbone with two new modules: an Extensional Perception Aggregator (EPA) that combines 3x3 convolutions with horizontal and vertical strip pooling to capture long-range spatial dependencies, and a Channel Semantic Weaving (CSW) layer that performs channel-wise recalibration. The method is evaluated on CUB-200-2011, where it reports 72.58% accuracy, and the paper claims a substantial performance gain over state-of-the-art methods. The claimed contributions are the two modules, the overall architecture, and the reported experimental result.

Significance. If the reported result were placed in the context of current fine-grained classification research, the paper would offer a plausible combination of existing ideas (strip pooling and channel attention) for bird classification. The architecture is clearly described and the motivation based on pose, scale, and background challenges is reasonable. However, the central claim is not supported by the manuscript's own comparative evidence: the reported 72.58% is far below current published results on CUB-200-2011, the comparison table is restricted to methods from 2013-2015, the face-cropping preprocessing is not specified, and no ablations or error bars are provided. The paper does not ship code, trained models, or a reproducibility appendix, so the positive potential of the module designs cannot be independently verified.

major comments (4)

[Abstract, Section 3.3, Table 2] The central claim that SASP achieves a 'substantial performance gain' on CUB-200-2011 is not supported because Table 2 compares only methods published between 2013 and 2015 with accuracies between 56.78% and 69.70%, and it omits the modern fine-grained classification literature, which reports accuracies above 90% on this same benchmark. With 72.58%, SASP is below current published state of the art, so the abstract and Conclusion assertions of 'significant performance improvements' are not established by the evidence presented in the manuscript.
[Section 3.2] The preprocessing step that 'facial regions are first automatically detected and cropped' is under-specified: no detector is named, no crop policy or crop size is given, and there is no evidence that the same preprocessing was applied to all compared methods. This makes the experimental setup irreproducible and potentially unfair to the baselines; the authors must describe the face detector, the cropping procedure, and an ablation of how the crop affects each method's accuracy.
[Section 3.3, Table 2] The empirical evaluation consists of a single reported accuracy value from one run, with no error bars, no seeds, no statistical significance testing, and no ablations of EPA, CSW, the individual strip-pooling branches, or the residual connection. Consequently, the 72.58% result cannot be attributed to the proposed modules, and the claim of 'architectural efficiency' is unsupported because no parameter count or FLOPs comparison is provided.
[Section 3.4, Figure 4] The analysis of the training loss curve attributes its shape to the EPA and CSW modules, but no controlled comparison with a baseline backbone is shown, and no validation accuracy curve is reported. The figure and discussion therefore do not constitute evidence for the modules' contributions.

minor comments (6)

[Section 2.3, Eq. (14)] In Eq. (14), the second substatement reuses the symbol W1 instead of introducing a distinct expansion matrix; please rename the second matrix (e.g., W2) so the bottleneck dimensions are expressed correctly as C/r then back to C.
[Section 4] The concluding sentence contains the redundant phrase 'significantly achieves significant performance improvements'; this should be rewritten for clarity.
[Abstract] The acknowledgment sentence 'Thanks for the support provided by MindSpore Community' appears in the abstract; it should be moved to the acknowledgments section.
[Section 3.3] The prose discussion compares SASP mainly with GoogLeNet, although Table 2 lists a stronger baseline, Xiao et al. at 69.70%; the text should discuss the strongest baseline in the table to make the comparison transparent.
[Section 3.2] It is unclear whether the facial-region cropping is applied at both training and test time, and whether the same crops are used for all baselines; please clarify the exact preprocessing pipeline for the training and test splits.
[Introduction and References] The related work is very brief, and several cited references (e.g., [1] and [2]) are not directly about fine-grained bird classification; adding a proper related-work section and citing more recent fine-grained classification methods would improve the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports a standard empirical training-and-testing result, and its central claims are not derived from their inputs by construction or via self-citation.

full rationale

The paper proposes SASP, a ResNet-50-based fine-grained bird classifier with two modules (EPA and CSW), trains it on CUB-200-2011, and reports a held-out test accuracy of 72.58%. The derivation chain is empirical rather than definitional: features are computed from convolutional, strip-pooling, and channel-reweighting operations (Eqs. 1-12), channel attention is a standard squeeze-and-excitation-style recalibration (Eqs. 13-15), and the classifier is trained with cross-entropy loss (Eq. 18). No parameter is fitted to the test set and then renamed as a prediction; no prediction is equivalent to an input by construction; and the paper does not rest on a self-citation chain or an imported uniqueness theorem to force its architecture. The main weaknesses—comparison against only 2013-2015 baselines, a 72.58% accuracy far below modern published results on CUB-200-2011, and unspecified face-detection preprocessing in Section 3.2—are serious correctness and reproducibility concerns, but they are not circularity. The load-bearing comparative claim depends on a non-representative baseline set, which is a benchmarking-selection issue, not a case where the result reduces to its own input. Therefore, the circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 2 invented entities

The central claim depends on standard empirical-learning assumptions: reliable dataset labels, a fair and reproducible preprocessing pipeline, and unspecified initialization. The two proposed modules are not validated independently and their components overlap prior published work, so the ledger records them as invented entities lacking independent evidence.

free parameters (3)

Initial learning rate and polynomial decay power = lr=0.1, power=0.5
Stated in Table 1; chosen by hand, affects convergence and final accuracy, not justified.
Training epochs, batch size, and weight decay = 70 epochs, batch size 256, weight decay 1e-4
Stated in Table 1; selected without explanation and directly influence the measured accuracy.
Channel reduction ratio r in CSW = not reported
Appears in Eq. (14) as the bottleneck dimension C/r; without its value the CSW module cannot be reproduced and the accuracy depends on it.

assumptions (3)

domain assumption CUB-200-2011 ground-truth labels and the standard train/test split are correct and reliable.
The reported accuracy is computed against these labels; no analysis of label noise or split usage is provided (Section 3.1).
ad hoc to paper The automatic facial-region detection and cropping described in Section 3.2 is a well-defined procedure that can be fairly applied to all compared methods.
The detector is not named, and it is not stated whether the baselines received the same crops, so the comparison in Table 2 rests on this unverified assumption.
domain assumption The ResNet-50 backbone is initialized in a standard way (e.g., ImageNet pretraining) and trained under the same protocol as the baselines.
The paper does not state whether pretrained weights are used; the absolute accuracy and the fairness of the comparison depend on this choice.

invented entities (2)

Extensional Perception Aggregator (EPA)
purpose: Captures long-range spatial dependencies via horizontal and vertical strip pooling combined with local convolution.
No ablation or external benchmark demonstrates that EPA alone improves accuracy; its components replicate existing strip pooling methods.
Channel Semantic Weaving (CSW)
purpose: Reweights feature channels via a compact descriptor and sigmoid attention.
CSW is structurally equivalent to the SE block; the paper provides no ablation or external validation for this specific variant.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SASP: Strip-Aware Spatial Perception for Fine-Grained Bird Image Classification." pith.science (2026). https://pith.science/paper/BNA45OZC

@misc{pith2026250524380,
  author       = {Pith},
  title        = {Pith review of: SASP: Strip-Aware Spatial Perception for Fine-Grained Bird Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BNA45OZC}},
  note         = {Machine review of arXiv:2505.24380}
}

read the original abstract

Fine-grained bird image classification (FBIC) is not only of great significance for ecological monitoring and species identification, but also holds broad research value in the fields of image recognition and fine-grained visual modeling. Compared with general image classification tasks, FBIC poses more formidable challenges: 1) the differences in species size and imaging distance result in the varying sizes of birds presented in the images; 2) complex natural habitats often introduce strong background interference; 3) and highly flexible poses such as flying, perching, or foraging result in substantial intra-class variability. These factors collectively make it difficult for traditional methods to stably extract discriminative features, thereby limiting the generalizability and interpretability of models in real-world applications. To address these challenges, this paper proposes a fine-grained bird classification framework based on strip-aware spatial perception, which aims to capture long-range spatial dependencies across entire rows or columns in bird images, thereby enhancing the model's robustness and interpretability. The proposed method incorporates two novel modules: extensional perception aggregator (EPA) and channel semantic weaving (CSW). Specifically, EPA integrates local texture details with global structural cues by aggregating information across horizontal and vertical spatial directions. CSW further refines the semantic representations by adaptively fusing long-range and short-range information along the channel dimension. Built upon a ResNet-50 backbone, the model enables jump-wise connection of extended structural features across the spatial domain. Experimental results on the CUB-200-2011 dataset demonstrate that our framework achieves significant performance improvements while maintaining architectural efficiency.

Figures

Figures reproduced from arXiv: 2505.24380 by the authors.

**Figure 1.** Challenges exist in FBIC. Despite substantial advances in FBIC, achieving high-precision recognition remains hindered by three principal challenges ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Overview of the proposed SASP architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Samples in CUB-200-2011 dataset. 3.2 Implementation Details In our experimental protocol, facial regions are first automatically detected and cropped to ensure fair comparison, after which all crops are uniformly resized to a resolution of 224 × 224. The model is trained for 70 epochs with a batch size of 256. We employ a Momentum optimizer initialized with a learning rate of 0.1, which is subsequently reduced to 0.… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Training loss over 70 epochs. 3.4 Visualization As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

11 extracted references · 11 canonical work pages

[1]

Deep Modeling and Optimization of Medical Image Classification

Yihang Wu, Muhammad Owais, Reem Kateb, and Ahmad Chaddad. Deep Modeling and Optimization of Medical Image Classification. In2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI), pages 1–4. IEEE

work page
[2]

Modeling Beyond MOS: Quality Assessment Models Must Integrate Context, Reasoning, and Multimodality

Mohamed Amine Kerkouri, Marouane Tliba, Aladine Chetouani, Nour Aburaed, and Alessandro Bruno. Modeling Beyond MOS: Quality Assessment Models Must Integrate Context, Reasoning, and Multimodality. 7

work page
[3]

Tobias, Ça ˘gan H

Joseph A. Tobias, Ça ˘gan H. ¸ Sekercio˘glu, and F. Hernan Vargas. Bird conservation in tropical ecosystems: Challenges and opportunities. pages 258–276

work page
[4]

The caltech-ucsd birds-200- 2011 dataset

Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200- 2011 dataset

work page 2011
[5]

Berg and P

T. Berg and P. N. Belhumeur. POOF: Part-Based One-vs.-One Features for Fine-Grained Categorization, Face Verification, and Attribute Estimation. In2013 IEEE Conference on Computer Vision and Pattern Recognition, pages 955–962

work page
[6]

Göering, E

C. Göering, E. Rodner, A. Freytag, and J. Denzler. Nonparametric Part Transfer for Fine-Grained Recognition. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 2489–2496

work page 2014
[7]

Y . Chai, V . Lempitsky, and A. Zisserman. Symbiotic Segmentation and Part Localization for Fine-Grained Categorization. In2013 IEEE International Conference on Computer Vision, pages 321–328

work page
[8]

Part detector discovery in deep convolutional neural networks

Marcel Simon, Erik Rodner, and Joachim Denzler. Part detector discovery in deep convolutional neural networks. InAsian Conference on Computer Vision, pages 162–177. Springer

work page

Show all 11 references

[9]

Decaf: A deep convolutional activation feature for generic visual recognition

Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng, and Trevor Darrell. Decaf: A deep convolutional activation feature for generic visual recognition. InInternational Conference on Machine Learning, pages 647–655. PMLR
[10]

The application of two-level attention models in deep convolutional neural network for fine-grained image classification

Tianjun Xiao, Yichong Xu, Kuiyuan Yang, Jiaxing Zhang, Yuxin Peng, and Zheng Zhang. The application of two-level attention models in deep convolutional neural network for fine-grained image classification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Re...
[11]

Szegedy, Wei Liu, Yangqing Jia, P

C. Szegedy, Wei Liu, Yangqing Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1–9. 8

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Deep Modeling and Optimization of Medical Image Classification

Yihang Wu, Muhammad Owais, Reem Kateb, and Ahmad Chaddad. Deep Modeling and Optimization of Medical Image Classification. In2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI), pages 1–4. IEEE

work page

[2] [2]

Modeling Beyond MOS: Quality Assessment Models Must Integrate Context, Reasoning, and Multimodality

Mohamed Amine Kerkouri, Marouane Tliba, Aladine Chetouani, Nour Aburaed, and Alessandro Bruno. Modeling Beyond MOS: Quality Assessment Models Must Integrate Context, Reasoning, and Multimodality. 7

work page

[3] [3]

Tobias, Ça ˘gan H

Joseph A. Tobias, Ça ˘gan H. ¸ Sekercio˘glu, and F. Hernan Vargas. Bird conservation in tropical ecosystems: Challenges and opportunities. pages 258–276

work page

[4] [4]

The caltech-ucsd birds-200- 2011 dataset

Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200- 2011 dataset

work page 2011

[5] [5]

Berg and P

T. Berg and P. N. Belhumeur. POOF: Part-Based One-vs.-One Features for Fine-Grained Categorization, Face Verification, and Attribute Estimation. In2013 IEEE Conference on Computer Vision and Pattern Recognition, pages 955–962

work page

[6] [6]

Göering, E

C. Göering, E. Rodner, A. Freytag, and J. Denzler. Nonparametric Part Transfer for Fine-Grained Recognition. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 2489–2496

work page 2014

[7] [7]

Y . Chai, V . Lempitsky, and A. Zisserman. Symbiotic Segmentation and Part Localization for Fine-Grained Categorization. In2013 IEEE International Conference on Computer Vision, pages 321–328

work page

[8] [8]

Part detector discovery in deep convolutional neural networks

Marcel Simon, Erik Rodner, and Joachim Denzler. Part detector discovery in deep convolutional neural networks. InAsian Conference on Computer Vision, pages 162–177. Springer

work page