Pith. sign in

REVIEW 3 major objections 4 minor 21 references

Spatially-Aware Class-Agnostic Object Counting

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read UpCount: reference-free class-agnostic counting that preserves spatial structure, achieving 12.39 MAE on FSC-147.

desk verdict Useful reference-free counting results, but the headline mechanism (DPT+FeatUp) is never isolated from MAE pretraining and the DAVE head, and the CARPK claim is overstated. read the letter →

arxiv 2607.16826 v2 pith:XAQHXPGV submitted 2026-07-18 cs.CV

Robert Wijaya , Md. Tanvir Hossain , Amanda Kau , Ngai-Man Cheung This is my paper
classification cs.CV
keywords class-agnosticobjectcountingexemplar-freespatialfeaturerefinementViTDPTFeatUpdensitymapMAEpretraining
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 tries to establish that class-agnostic object counting fails on structurally complex objects because visual features lose spatial detail, and that strengthening the representation early—by extracting multi-layer ViT features and refining them with DPT and FeatUp—lets a model count arbitrary objects with no exemplars or text prompts. It introduces UpCount, which uses a proposal–verification head to detect repeated patterns and outputs a density map. On FSC-147, UpCount reports 12.39 MAE and 100.89 RMSE on the test set, and transfers to vehicle counting on CARPK with 6.27 MAE and 8.79 RMSE, outperforming reference-free baselines. A sympathetic reader would care because it moves toward fully automatic counting without per-class annotation effort.

What carries the argument

The central object is the spatially refined multi-scale pyramid F_ref: multi-layer ViT-B/16 features are reassembled by DPT into a 96x96 pyramid and refined with FeatUp joint bilateral upsampling conditioned on RGB guidance, then fused with RGB detail. This carries the argument by preserving structural boundaries that patch tokenization loses; the proposal–verification head then matches repeated patterns and calibrates counts.

What would settle it

Train UpCount with the DPT pyramid and FeatUp refinement removed—feeding flattened multi-layer features through a simple upsampling—and retain all other components. If the MAE/RMSE on FSC-147 stays within a small margin, the paper's central attribution to spatial refinement collapses.

Watch

Extended reading notes

Core claim

The central claim is that the bottleneck for generalised counting is spatial resolution of features, not counting-head design. By taking layer-normalized features from four ViT blocks (2, 5, 8, 11), reassembling them into a DPT-style multi-scale pyramid, and refining with FeatUp's joint bilateral upsampling against an RGB guidance, UpCount produces features that keep fine structure alongside semantics. The proposal–verification head then converts these into verified density maps. The paper claims this yields state-of-the-art reference-free results on FSC-147 and strong transfer to CARPK, showing spatial refinement is sufficient.

Load-bearing premise

The load-bearing premise is that the DPT+FeatUp spatial refinement, not the MAE pretraining or the proposal–verification head, is what causes the reported accuracy; the paper provides no ablation to isolate this component.

Editorial extensions

If this is right

  • If correct, reference-free counting can match or beat exemplar-guided methods without needing bounding boxes or text, reducing annotation cost.
  • Counting of structurally complex objects (e.g., sunglasses) should no longer double-count subparts, because spatial structure is preserved.
  • The learned representation transfers to a distinct domain (vehicle counting) after fine-tuning, suggesting the spatial refinement generalizes across categories.
  • The two-stage training (MAE pretraining on FSC-147, then density-map fine-tuning) is a viable recipe for class-agnostic counting.

Reading between the lines

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

  • The paper attributes gains to DPT+FeatUp spatial refinement, but because no ablation isolates that module, the improvement could instead come from the MAE pretraining on FSC-147 or the proposal–verification head; an ablation would settle this.
  • The same spatial-refinement recipe could be dropped into other density-based counters (crowd counting, cell counting) to see if structure preservation helps beyond class-agnostic setting.
  • A testable extension: evaluate UpCount on images with heavy occlusion or overlapping instances; spatial refinement should help most there, but if not, the mechanism is not what the paper thinks.
  • The MAE pretraining on only FSC-147 training split may be a strong contributor; one could test by initializing from ImageNet-MAE only and comparing.
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.

Referee Report

3 major / 4 minor

Summary. The paper proposes UpCount, a class-agnostic object counter that requires no exemplar boxes or text prompts at inference. The architecture extracts multi-layer features from an MAE-pretrained ViT-B/16, reassembles them with a DPT-style pyramid, refines them with FeatUp joint bilateral upsampling, and feeds the result to a proposal-verification counting head. The model is trained in two stages: masked autoencoder pretraining on the FSC-147 training split, followed by supervised counting fine-tuning on FSC-147 or CARPK. On FSC-147 the authors report 12.39 MAE / 100.89 RMSE on the test set, and on CARPK 6.27 MAE / 8.79 RMSE, outperforming reference-free CounTR, LOCA, and DAVE.

Significance. If the claim that DPT+FeatUp spatial refinement improves counting on structurally complex objects is substantiated, UpCount would be a useful reference-free counting method. The paper is clearly written, the architecture is plausible, and the code is promised. However, the evidence for the central causal claim is currently missing: the full system improves over baselines, but no ablation isolates the DPT+FeatUp module from other components such as the MAE pretraining on the target dataset or the proposal-verification head. The contribution is therefore not yet established to the standard expected for a venue where causal attribution is a core part of the claim.

major comments (3)
  1. [Section III-B.1, Table I] The headline result is attributed to DPT+FeatUp spatial refinement, but the system differs from reference-free baselines in several components simultaneously: (1) MAE pretraining on the FSC-147 training split, (2) multi-layer feature extraction, (3) DPT pyramid reassembly, (4) FeatUp JBU, (5) a proposal-verification head following DAVE, and (6) a multi-term loss. No ablation removes DPT+FeatUp or replaces it with a simple bilinear upsampling. Without such an ablation, the reported 12.39 test MAE cannot be causally assigned to the spatial refinement module. Please provide an ablation study that isolates (i) MAE pretraining on FSC-147, (ii) DPT+FeatUp, and (iii) the proposal-verification head, including a variant where the DPT+FeatUp module is replaced by bilinear upsampling of the final-block features.
  2. [Section I and Section IV-C] The motivation is that recent methods fail on structurally complex objects, e.g., sunglasses being double-counted. However, the evaluation reports only aggregate MAE/RMSE on the full FSC-147 test set, with no per-class or per-category breakdown, and no targeted evaluation on structurally complex categories. The claim that spatial refinement improves structural counting therefore lacks direct evidence. Please report evaluation on structurally complex object categories (or a difficulty-ranked subset) and/or failure-case analysis.
  3. [Section III-B.1, Section IV-C] The two-stage training scheme first performs masked autoencoder pretraining on the FSC-147 training split. This is a target-domain self-supervised adaptation step that no reference-free baseline in Table I performs. Consequently, the performance gap over these baselines could be due to domain adaptation rather than the proposed spatial refinement. Isolate the effect of this pretraining step by comparing with (a) ImageNet-MAE pretraining only, and (b) no MAE pretraining at all, while keeping the rest of the architecture fixed.
minor comments (4)
  1. [Section IV-B] The effective learning rate for FSC-147 is reported as 'approximately 2.03×10^5'. Given the stated base learning rate and batch size, the intended value is almost certainly 2.03×10^{-5} (linear scaling by batch size/256). Please correct the typo and clarify the scaling rule.
  2. [Throughout] The model name is inconsistent: 'UpCount' and 'UPCount' are both used. Please standardize.
  3. [Section III-A.4, Eq. (11)] The learned spatial correction map A_i and the global count-calibration factor c_i are not defined or motivated. Specify their dimensions, initialization, and how they are supervised (e.g., are they jointly learned with the other heads?).
  4. [Section IV-C] The paper reports single-run numbers with no error bars or statistical significance testing. Given the moderate performance gaps over some baselines (e.g., 12.39 vs. 14.12 test MAE), reporting multiple runs or at least a standard deviation would strengthen the claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: UpCount is a standard supervised counting pipeline evaluated on held-out test splits.

full rationale

UpCount's count is the integral of a learned density map (Eq. 11) supervised by ground-truth density maps (Eq. 13) on the FSC-147 and CARPK training splits and evaluated on the official test splits. The MAE pretraining step (Eq. 12) uses unlabeled training images from the same dataset; this is a standard self-supervised initialization, not a fitted parameter renamed as a prediction, and it does not use test labels. No equation in the paper defines a quantity in terms of the target it is claimed to predict, and no parameter fitted to test data is reported as a prediction. The references to DPT, FeatUp, and DAVE are external prior work and are not self-citations by the authors; no load-bearing claim rests on a self-citation. The absence of an ablation isolating DPT+FeatUp is a limitation of empirical attribution, not a circularity: it does not make any reported number equal to its own input by construction. Evaluation against external benchmarks (FSC-147, CARPK) keeps the derivation self-contained relative to circularity concerns.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new physical or conceptual entities are introduced. The free parameters are the usual unstated ML hyperparameters (loss weights, top-K, learning rate, foreground weighting) that the central result depends on; the axioms are the modeling assumptions that repeated-pattern detection and spatially refined ViT features suffice for class-agnostic counting on the chosen benchmarks.

free parameters (4)
  • Counting loss weights lambda_c, lambda_log, lambda_v = not reported
    Eq. (13) defines a weighted sum of density, count, log-count, and verification losses; the weights are not specified and are presumably tuned on validation data.
  • Top-K nearest-neighbor matches K in proposal-verification = not reported
    Section III-A4 says the module aggregates top-K matches to identify repeated patterns, but K is never given.
  • Effective learning rate for FSC-147 fine-tuning = reported as 2.03x10^5, likely 2.03x10^-5
    Section IV-B contains an apparent sign error; batch size 26 and base LR 2e-4 give ~5.2e-3 (or ~2.03e-5 under a linear scaling rule), not 2.03e5. The printed value is physically unreasonable.
  • Foreground weighting in density regression loss = not reported
    L_density is described as foreground-weighted, but the weighting formula or parameter is not given.
assumptions (4)
  • domain assumption Counting arbitrary objects can be reduced to detecting repeated visual patterns and regressing a density map
    Underlies the proposal-verification head; repeated-pattern detection is assumed sufficient for class-agnostic counting.
  • domain assumption MAE-pretrained ViT features, after DPT/FeatUp refinement, preserve spatial structure of small/complex objects
    Core design premise in Sections III-A2/3; no direct experiment verifies that spatial refinement is the cause of gains.
  • domain assumption FSC-147 and CARPK are representative benchmarks whose results generalize to real-world class-agnostic counting
    Conclusions are drawn from two benchmarks; no cross-domain or difficult-structure test such as the sunglasses case in the introduction.
  • domain assumption No information leak from FSC-147 test set into the MAE pretraining
    MAE pretraining uses only the FSC-147 training split, per Section IV-B; the paper does not discuss class or image overlap with the test split.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatially-Aware Class-Agnostic Object Counting." pith.science (2026). https://pith.science/paper/XAQHXPGV

@misc{pith2026260716826,
  author       = {Pith},
  title        = {Pith review of: Spatially-Aware Class-Agnostic Object Counting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XAQHXPGV}},
  note         = {Machine review of arXiv:2607.16826}
}
read the original abstract

Generalised object counting aims to estimate the number of instances of an arbitrary object category from a single image, but many recent methods can struggle on structurally complex objects due to limited spatial modelling. We present \textit{UpCount}, a class-agnostic counter designed to better preserve spatial structure. UpCount strengthens the visual representation by extracting multi-layer features from a ViT-B/16 encoder and reassembling them into a refined multi-scale pyramid that is spatially refined using Dense Prediction Transformers and FeatUp, yielding features with improved structural and spatial sensitivity; a proposal--verification counting head then identifies repeated patterns and produces a density map for the final count. On FSC-147, UpCount achieves 12.39 MAE and 100.89 RMSE on the test set, and it transfers effectively to vehicle counting on CARPK (6.27 MAE, 8.79 RMSE). Code: https://github.com/r28112072-rgb/upcount

Figures

Figures reproduced from arXiv: 2607.16826 by the authors.

Figure 1
Figure 1. Overview of the proposed UpCount architecture. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Qualitative results of UpCount on FSC-147 and CARPK. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 3 linked inside Pith

  1. [1]

    Single-image crowd counting via multi-column convolutional neural network,

    Y . Zhang, D. Zhou, S. Chen, S. Gao, and Y . Ma, “Single-image crowd counting via multi-column convolutional neural network,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2016

    2016

  2. [2]

    A large contextual dataset for classification, detection and counting of cars with deep learning,

    T. N. Mundhenk, G. Konjevod, W. A. Sakla, and K. Boakye, “A large contextual dataset for classification, detection and counting of cars with deep learning,” inComputer Vision–ECCV 2016: 14th European Confer- ence, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14. Springer, 2016, pp. 785–800

    2016

  3. [3]

    Countr: Transformer- based generalised visual counting,

    C. Liu, Y . Zhong, A. Zisserman, and W. Xie, “Countr: Transformer- based generalised visual counting,”arXiv preprint arXiv:2208.13721, 2022

    arXiv 2022

  4. [4]

    Class-agnostic counting,

    E. Lu, W. Xie, and A. Zisserman, “Class-agnostic counting,” inCom- puter Vision–ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III

    2018

  5. [5]

    Vision transformers for dense prediction,

    R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision transformers for dense prediction,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 12 179–12 188

    2021

  6. [6]

    Featup: A model-agnostic framework for features at any resolution,

    S. Fu, M. Hamilton, L. E. Brandt, A. Feldmann, Z. Zhang, and W. Freeman, “Featup: A model-agnostic framework for features at any resolution,” inInternational Conference on Learning Representations, vol. 2024, 2024, pp. 45 324–45 350

    2024

  7. [7]

    Learning to count everything,

    V . Ranjan, U. Sharma, T. Nguyen, and M. Hoai, “Learning to count everything,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 3394–3403

    2021

  8. [8]

    Zero-shot object counting,

    J. Xu, H. Le, V . Nguyen, V . Ranjan, and D. Samaras, “Zero-shot object counting,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 15 548–15 557

    2023

Show all 21 references
  1. [9]

    Exemplar free class agnostic counting,

    V . Ranjan and M. H. Nguyen, “Exemplar free class agnostic counting,” inProceedings of the Asian Conference on Computer Vision, 2022, pp. 3121–3137

    2022

  2. [10]

    Clip-count: Towards text-guided zero- shot object counting,

    R. Jiang, L. Liu, and C. Chen, “Clip-count: Towards text-guided zero- shot object counting,”arXiv preprint arXiv:2305.07304, 2023

    2023 arXiv

  3. [11]

    Open- world text-specified object counting,

    N. Amini-Naieni, K. Amini-Naieni, T. Han, and A. Zisserman, “Open- world text-specified object counting,” 2023

    2023

  4. [12]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PMLR, 2021, pp. 8748–8763

    2021

  5. [13]

    Drone-based object counting by spatially regularized regional proposal network,

    M.-R. Hsieh, Y .-L. Lin, and W. H. Hsu, “Drone-based object counting by spatially regularized regional proposal network,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 4145– 4153

    2017

  6. [14]

    Springer, 2019, pp. 669–684

    2019

  7. [15]

    T2icount: Enhancing cross-modal understanding for zero-shot counting,

    Y . Qian, Z. Guo, B. Deng, C. T. Lei, S. Zhao, C. P. Lau, X. Hong, and M. P. Pound, “T2icount: Enhancing cross-modal understanding for zero-shot counting,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 25 336–25 345

    2025

  8. [16]

    Countse: Soft exemplar open- set object counting,

    S. Liu, P. Zhang, S. Zhang, and W. Ke, “Countse: Soft exemplar open- set object counting,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 21 536–21 546

    2025

  9. [17]

    Boosting quantitive and spatial awareness for zero-shot object counting,

    D. Zhang, B. Li, F. Wang, Z. Zhao, and J. Gao, “Boosting quantitive and spatial awareness for zero-shot object counting,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 20 356–20 366

    2026

  10. [18]

    A low-shot object counting network with iterative prototype adaptation,

    N. Djukic, A. Lukezic, V . Zavrtanik, and M. Kristan, “A low-shot object counting network with iterative prototype adaptation,”arXiv preprint arXiv:2211.08217, 2022

    2022 arXiv

  11. [19]

    Dave-a detect-and-verify paradigm for low-shot counting,

    J. Pelhan, V . Zavrtanik, M. Kristanet al., “Dave-a detect-and-verify paradigm for low-shot counting,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 23 293–23 302

    2024

  12. [20]

    Countgd: Multi-modal open-world counting,

    N. Amini-Naieni, T. Han, and A. Zisserman, “Countgd: Multi-modal open-world counting,”Advances in Neural Information Processing Sys- tems, vol. 37, pp. 48 810–48 837, 2024

    2024

  13. [21]

    Countingdino: A training-free pipeline for class-agnostic counting using unsupervised backbones,

    G. Pacini, L. Bianchi, L. Ciampi, N. Messina, G. Amato, and F. Falchi, “Countingdino: A training-free pipeline for class-agnostic counting using unsupervised backbones,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2026, pp. 806–815

    2026

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.