pith. sign in

arxiv: 2605.18063 · v1 · pith:NKS4DE5Bnew · submitted 2026-05-18 · 💻 cs.CV · cs.LG

The MixCount Dataset: Bridging the Data Gap for Open-Vocabulary Object Counting

Corentin Dumery , Niki Amini-Naieni , Shervin Naini , Pascal Fua This is my paper

Pith reviewed 2026-05-20 11:19 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords object countingsynthetic datasetmixed objectsopen-vocabularyautomatic annotationcomputer visiontransfer learningbenchmark
0
0 comments X

The pith

Automatically synthesized mixed-object scenes improve counting model accuracy on real images by more than 18 percent.

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

The paper claims that object counting models fail in everyday scenes with several object types because real training data is scarce and noisy while existing synthetic sets lack variety. The authors respond by building an automatic pipeline that creates large numbers of images containing mixed objects together with exact pixel-level counts and detailed text descriptions. When models are trained on this new data they perform noticeably better on actual photographs, which matters because reliable counting supports tasks like factory inspection and inventory management without the expense of manual labeling.

Core claim

MixCount is a dataset and benchmark for mixed-object counting created through an automatic generation pipeline that synthesizes images, fine-grained textual descriptions, and pixel-perfect counting annotations at scale. Evaluating state-of-the-art counting models on MixCount exposes severe degradation in the mixed-object setting. Training these models on the synthesized data yields substantial gains on real-world benchmarks, reducing MAE by 20.14 percent on FSC-147 and by 18.3 percent on PairTally. These results establish MixCount as both a benchmark and a training dataset for fine-grained counting while demonstrating that the pipeline, which produces effectively unlimited labeled data, can

What carries the argument

The automatic generation pipeline that produces synthetic images of mixed objects along with pixel-perfect counting annotations and fine-grained textual descriptions.

If this is right

  • Current counting models exhibit clear performance drops when tested on mixed-object scenes from the new benchmark.
  • Training on the synthesized data transfers to real datasets and lowers error rates without requiring manual annotation effort.
  • The pipeline can generate unlimited perfectly labeled examples to overcome data bottlenecks in counting research.
  • The approach supports open-vocabulary counting by pairing images with detailed textual descriptions of object types.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same automatic synthesis method could be adapted to generate training data for related tasks such as object detection in cluttered scenes.
  • Industrial applications that need to count diverse items in one view might adopt the pipeline to create custom training sets on demand.
  • Future models could integrate the generation process directly into training loops to produce fresh examples tailored to observed failure cases.

Load-bearing premise

The synthetic images and annotations match the distribution and complexity of real mixed-object scenes closely enough that training on them improves performance on actual photographs without harmful domain artifacts.

What would settle it

Retraining the same counting models on MixCount and measuring no reduction or an increase in MAE on FSC-147 or PairTally would show that the synthetic data does not transfer effectively.

Figures

Figures reproduced from arXiv: 2605.18063 by Corentin Dumery, Niki Amini-Naieni, Pascal Fua, Shervin Naini.

Figure 1
Figure 1. Figure 1: The MixCount Dataset. Our dataset includes images of mixed objects in settings that are challenging for visual counting models, along with precise and rich ground truth annotations, text descriptions and visual exemplars. Abstract Object counting is a foundational vision task with over a decade of dedicated research, yet state-of-the-art models still fail systematically in the mixed-object setting that dom… view at source ↗
Figure 2
Figure 2. Figure 2: Bridging the data gap. Visual counting models often struggle to distinguish similar objects (a), fail to recognize self-similar components as a single entity (b), and are easily distracted by repetitive background patterns (c). We show that these failure modes can be substantially reduced with targeted training data, and design MixCount to bridge this gap, resulting in -20% error on recent benchmarks. Prom… view at source ↗
Figure 3
Figure 3. Figure 3: Large-scale synthetic data comparison. (a) MCAC [23] and (b) SITUATE [42] samples lack the complexity of real-world images. In comparison, (c) MixCount samples are more diverse, realistic, and challenging. bounding boxes over example instances. State-of-the-art models CountGD [2] and CountGD++ [4] accept both visual exemplars and text as prompts. By accepting prompts as extra inputs, these models adapt to … view at source ↗
Figure 4
Figure 4. Figure 4: Dataset features. MixCount includes different exemplars and tiered description granularity. 3 The MixCount dataset and data generator Our work is built upon the observation that existing visual counting datasets are either limited in diversity and scale or lack realism. Large-scale datasets such as the ones in [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Dense ground-truth annotations. Each sample is annotated with object localization labels, as well as depth and normal maps. We also introduce two different kinds of visual exemplars. An internal exemplar, represented as a bounding box in the image, and an external exemplar, which is an image of the same object but in a different setting. MixCount includes both internal and external exemplars. Internal exem… view at source ↗
Figure 6
Figure 6. Figure 6: Data generator. Our generator samples objects, distractors, environment and camera placement to procedurally generate photorealistic training samples. All assets are issued from high￾quality captures of real-world objects. of real-world objects. We manually inspect the dataset to filter out objects that are too similar to be distinguished by text or visual exemplar. After the first object class is created,… view at source ↗
Figure 7
Figure 7. Figure 7: Counting MAE vs. exemplar rank. Counting MAE increases as lower-scoring exemplars are used, indicating that exemplar quality affects counting performance. Lower exemplar rank corresponds to higher exemplar score. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Dataset statistics. Along with this paper, we release the first version of the MixCount dataset. This release includes 58,000 scenes with over 4 million objects to count at an average of 67 per scene. 1522 different objects can be found in the dataset, all of which were manually filtered from DTC [13] to remove duplicates and indistinguishable objects. We display additional histogram statistics in [PITH_F… view at source ↗
Figure 9
Figure 9. Figure 9: Additional test samples. We display test samples from the MixCount dataset along with model predictions. We generate predictions with our best model, the CountGD++ model (+MixCount) with positive and negative prompts. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: External exemplars. Additional external exemplars are provided by rendering a canonical view of each object in front of a white background. These exemplars are used to evaluate the sensitivity of counting models to the context of the provided visual exemplar. When a container exists, spawn positions are sampled above an ellipse inscribed in the container axis-aligned bounding box (AABB). With spawn margin… view at source ↗
read the original abstract

Object counting is a foundational vision task with over a decade of dedicated research, yet state-of-the-art models still fail systematically in the mixed-object setting that dominates real-world applications such as industrial inspection and product sorting. We show that this gap is strongly driven by limitations in existing training and evaluation data: real counting datasets are prohibitively expensive to annotate and suffer from labeling noise, while existing synthetic alternatives lack diversity and realism. We address this with MixCount, a dataset and benchmark for mixed-object counting designed to target the failure modes of current counting models. To overcome the high cost of constructing and labeling such data, we develop an automatic generation pipeline that synthesizes images, fine-grained textual descriptions, and pixel-perfect counting annotations at scale, eliminating the labeling ambiguity that plagues prior datasets. Evaluating state-of-the-art counting models on MixCount exposes severe degradation in the mixed-object setting. More importantly, training these models on our synthesized data yields substantial gains on real-world benchmarks, reducing MAE by 20.14% on FSC-147 and by 18.3% on PairTally. These results establish MixCount as both a benchmark and a training dataset for fine-grained counting, and demonstrate that our pipeline, which produces effectively unlimited labeled data, helps address a long-standing bottleneck in counting models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces the MixCount dataset and benchmark for mixed-object counting, generated via an automatic synthesis pipeline that produces images, fine-grained textual descriptions, and pixel-perfect annotations at scale. It demonstrates that state-of-the-art counting models degrade substantially in the mixed-object setting on this data and reports that training on MixCount yields MAE reductions of 20.14% on FSC-147 and 18.3% on PairTally.

Significance. If the performance gains are shown to stem from improved distribution matching rather than data volume or artifacts, the work would meaningfully address the long-standing data bottleneck in object counting for complex real-world scenes. The automatic pipeline's capacity to generate effectively unlimited labeled data without annotation noise is a practical strength that could support further progress in open-vocabulary and fine-grained counting.

major comments (2)
  1. The central empirical claim (MAE reductions of 20.14% on FSC-147 and 18.3% on PairTally after training on MixCount) is load-bearing for the assertion that the dataset bridges the data gap. The manuscript provides no ablation that holds total training sample count fixed while varying only the realism of the generation process, and no quantitative domain-similarity metrics (e.g., feature-space distances or perceptual studies) between the synthetic images and the real benchmarks. This leaves the causal attribution to faithful distribution matching open to alternative explanations such as volume effects or pipeline-specific cues.
  2. In the baseline evaluation and training sections, insufficient detail is given on the original models' training protocols (e.g., from-scratch retraining versus fine-tuning), the precise data splits, hyperparameter settings, and statistical significance of the reported MAE gains. These omissions hinder assessment of whether the improvements are robust and fairly compared.
minor comments (1)
  1. The abstract and title use both 'mixed-object counting' and 'open-vocabulary object counting'; a brief clarification of their relationship would improve precision.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive comments on our manuscript. We address each of the major comments below and outline the revisions we will make to strengthen the paper.

read point-by-point responses

  1. Referee: The central empirical claim (MAE reductions of 20.14% on FSC-147 and 18.3% on PairTally after training on MixCount) is load-bearing for the assertion that the dataset bridges the data gap. The manuscript provides no ablation that holds total training sample count fixed while varying only the realism of the generation process, and no quantitative domain-similarity metrics (e.g., feature-space distances or perceptual studies) between the synthetic images and the real benchmarks. This leaves the causal attribution to faithful distribution matching open to alternative explanations such as volume effects or pipeline-specific cues.

    Authors: We appreciate the referee highlighting this important point regarding causal attribution. The manuscript reports substantial gains from training on MixCount but does not contain an ablation that holds training sample count fixed. We will add a controlled ablation in the revised manuscript comparing models trained on size-matched subsets of MixCount against the original real-data baselines. We will also add quantitative domain-similarity metrics, specifically Fréchet Inception Distance (FID) between MixCount images and the real benchmark distributions, to provide evidence supporting distribution matching over volume or artifact effects. revision: yes

  2. Referee: In the baseline evaluation and training sections, insufficient detail is given on the original models' training protocols (e.g., from-scratch retraining versus fine-tuning), the precise data splits, hyperparameter settings, and statistical significance of the reported MAE gains. These omissions hinder assessment of whether the improvements are robust and fairly compared.

    Authors: We agree that greater detail is required for reproducibility and fair evaluation. In the revised manuscript we will expand the relevant sections to specify whether each baseline was retrained from scratch or fine-tuned, the exact train/validation/test splits used, the complete set of hyperparameter values, and statistical significance of the MAE improvements (including standard deviations over multiple random seeds). revision: yes

Circularity Check

0 steps flagged

No significant circularity; results are empirical on external benchmarks

full rationale

The paper's core claims rest on an automatic synthesis pipeline for MixCount data followed by empirical training that produces measured MAE reductions on independent real-world test sets (FSC-147 and PairTally). These performance numbers are obtained via standard transfer-learning evaluation against externally held-out data and do not reduce to fitted parameters, self-definitions, or self-citation chains within the paper. No equations, uniqueness theorems, or ansatzes are presented that would create a self-referential loop; the reported gains are falsifiable against the cited external benchmarks and therefore constitute independent evidence rather than a renaming or reconstruction of the inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work relies on standard computer vision assumptions about object rendering and scene composition; no new free parameters or invented entities are introduced beyond the dataset itself.

axioms (1)
  • domain assumption Synthetic image generation can produce sufficiently realistic mixed-object scenes that transfer to real data distributions.
    Invoked when claiming that training on MixCount improves real-world benchmarks.

pith-pipeline@v0.9.0 · 5775 in / 1170 out tokens · 29963 ms · 2026-05-20T11:19:58.459836+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 62 canonical work pages · 3 internal anchors

  1. [1]

    Open-world text-specified object counting

    Niki Amini-Naieni, Kiana Amini-Naieni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. InThe 36th British Machine Vision Conference (BMVC), 2023

    work page 2023

  2. [2]

    Countgd: Multi-modal open-world counting.Advances in Neural Information Processing Systems, 37:48810–48837, 2024

    Niki Amini-Naieni, Tengda Han, and Andrew Zisserman. Countgd: Multi-modal open-world counting.Advances in Neural Information Processing Systems, 37:48810–48837, 2024

    work page 2024

  3. [3]

    Countgd++: Generalized prompting for open-world counting

    Niki Amini-Naieni and Andrew Zisserman. Countgd++: Generalized prompting for open-world counting. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2026

    work page 2026

  4. [4]

    Open-world object counting in videos

    Niki Amini-Naieni and Andrew Zisserman. Open-world object counting in videos. InPro- ceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 2300–2308, 2026

    work page 2026

  5. [5]

    Vqa: Visual question answering

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. InProceedings of the IEEE international conference on computer vision, pages 2425–2433, 2015

    work page 2015

  6. [6]

    Counting in the wild

    Carlos Arteta, Victor Lempitsky, and Andrew Zisserman. Counting in the wild. InEuropean conference on computer vision, pages 483–498. Springer, 2016

    work page 2016

  7. [7]

    Chain-of-look spatial reasoning for dense surgical instrument counting

    Rishikesh Bhyri, Brian R Quaranto, Junsong Yuan, Peter CW Kim, and Nan Xi. Chain-of-look spatial reasoning for dense surgical instrument counting. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8521–8530, 2026

    work page 2026

  8. [8]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018

    Blender Online Community.Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018

    work page 2018

  10. [10]

    Referring expression counting

    Siyang Dai, Jun Liu, and Ngai-Man Cheung. Referring expression counting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16985–16995, 2024

    work page 2024

  11. [11]

    Figo: Fine-grained object counting without annotations.arXiv preprint arXiv:2504.11705, 2025

    Adrian D’Alessandro, Ali Mahdavi-Amiri, and Ghassan Hamarneh. Figo: Fine-grained object counting without annotations.arXiv preprint arXiv:2504.11705, 2025

    work page arXiv 2025

  12. [12]

    Molmo and pixmo: Open weights and open data for state-of-the-art vision-language models

    Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art vision-language models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 91–104, 2025

    work page 2025

  13. [13]

    Digital twin catalog: A large-scale photorealistic 3d object digital twin dataset

    Zhao Dong, Ka Chen, Zhaoyang Lv, Hong-Xing Yu, Yunzhi Zhang, Cheng Zhang, Yufeng Zhu, Stephen Tian, Zhengqin Li, Geordie Moffatt, et al. Digital twin catalog: A large-scale photorealistic 3d object digital twin dataset. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 753–763, 2025

    work page 2025

  14. [14]

    Stackcounting dataset: 3d stacked objects with ground-truth count and geometry, 2025

    Corentin Dumery, Noa Etté, and Adriano D’Alessandro. Stackcounting dataset: 3d stacked objects with ground-truth count and geometry, 2025

    work page 2025

  15. [15]

    Counting stacked objects

    Corentin Dumery, Noa Etté, Aoxiang Fan, Ren Li, Jingyi Xu, Hieu Le, and Pascal Fua. Counting stacked objects. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 19774–19783, 2025

    work page 2025

  16. [16]

    Automated counting of stacked objects in industrial inspection.arXiv preprint arXiv:2603.15470, 2026

    Corentin Dumery, Noa Etté, Aoxiang Fan, Ren Li, Jingyi Xu, Hieu Le, and Pascal Fua. Automated counting of stacked objects in industrial inspection.arXiv preprint arXiv:2603.15470, 2026

    work page arXiv 2026

  17. [17]

    Afreeca: Annotation-free counting for all

    Adriano D’Alessandro, Ali Mahdavi-Amiri, and Ghassan Hamarneh. Afreeca: Annotation-free counting for all. InEuropean Conference on Computer Vision, pages 75–91. Springer, 2025. 10

    work page 2025

  18. [18]

    Vastextures: Vast repository of textures and pbr materials extracted from real-world images using unsupervised methods.arXiv preprint arXiv:2406.17146, 2024

    Sagi Eppel. Vastextures: Vast repository of textures and pbr materials extracted from real-world images using unsupervised methods.arXiv preprint arXiv:2406.17146, 2024

    work page arXiv 2024

  19. [19]

    Flaccavento, Victor Lempitsky, Iestyn Pope, P

    G. Flaccavento, Victor Lempitsky, Iestyn Pope, P. R. Barber, Andrew Zisserman, J. Alison Noble, and B. V ojnovic. Learning to count cells: applications to lens-free imaging of large fields. InMicroscopic Image Analysis with Applications in Biology, 2011

    work page 2011

  20. [20]

    Learning to predict indoor illumination from a single image.ACM Transactions on Graphics (TOG), 36(6):1–14, 2017

    Marc-André Gardner, Kalyan Sunkavalli, Ersin Yumer, Xiaohui Shen, Emiliano Gambaretto, Christian Gagné, and Jean-François Lalonde. Learning to predict indoor illumination from a single image.ACM Transactions on Graphics (TOG), 36(6):1–14, 2017

    work page 2017

  21. [21]

    Kubric: A scalable dataset generator

    Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749–3761, 2022

    work page 2022

  22. [22]

    Sau-net: A unified network for cell counting in 2d and 3d microscopy images.IEEE/ACM transactions on computational biology and bioinformatics, 19(4):1920–1932, 2021

    Yue Guo, Oleh Krupa, Jason Stein, Guorong Wu, and Ashok Krishnamurthy. Sau-net: A unified network for cell counting in 2d and 3d microscopy images.IEEE/ACM transactions on computational biology and bioinformatics, 19(4):1920–1932, 2021

    work page 1920

  23. [23]

    Abc easy as 123: A blind counter for exemplar-free multi- class class-agnostic counting

    Michael Hobley and Victor Prisacariu. Abc easy as 123: A blind counter for exemplar-free multi- class class-agnostic counting. InEuropean Conference on Computer Vision, pages 304–319. Springer, 2024

    work page 2024

  24. [24]

    Countex: Fine-grained counting via exemplars and exclusion.arXiv preprint arXiv:2602.19432, 2026

    Yifeng Huang, Gia Khanh Nguyen, and Minh Hoai. Countex: Fine-grained counting via exemplars and exclusion.arXiv preprint arXiv:2602.19432, 2026

    work page arXiv 2026

  25. [25]

    Bcdata: A large-scale dataset and benchmark for cell detection and counting

    Zhongyi Huang, Yao Ding, Guoli Song, Lin Wang, Ruizhe Geng, Hongliang He, Shan Du, Xia Liu, Yonghong Tian, Yongsheng Liang, et al. Bcdata: A large-scale dataset and benchmark for cell detection and counting. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 289–298. Springer, 2020

    work page 2020

  26. [26]

    Countnet3d: A 3d computer vision approach to infer counts of occluded objects

    Porter Jenkins, Kyle Armstrong, Stephen Nelson, Siddhesh Gotad, J Stockton Jenkins, Wade Wilkey, and Tanner Watts. Countnet3d: A 3d computer vision approach to infer counts of occluded objects. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3008–3017, 2023

    work page 2023

  27. [27]

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

    Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clip-count: Towards text-guided zero-shot object counting. InProceedings of the 31st ACM International Conference on Multimedia, pages 4535–4545, 2023

    work page 2023

  28. [28]

    Evaluating numerical reasoning in text-to-image models.Advances in neural information processing systems, 38, 2024

    Ivana Kaji´c, Olivia Wiles, Isabela Albuquerque, Matthias Bauer, Su Wang, Jordi Pont-Tuset, and Aida Nematzadeh. Evaluating numerical reasoning in text-to-image models.Advances in neural information processing systems, 38, 2024

    work page 2024

  29. [29]

    The caltech fish counting dataset: A benchmark for multiple-object tracking and counting

    Justin Kay, Peter Kulits, Suzanne Stathatos, Siqi Deng, Erik Young, Sara Beery, Grant Van Horn, and Pietro Perona. The caltech fish counting dataset: A benchmark for multiple-object tracking and counting. InEuropean Conference on Computer Vision, pages 290–311. Springer, 2022

    work page 2022

  30. [30]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, George Drettakis, et al. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1, 2023

    work page 2023

  31. [31]

    Automated object counting for visual inspection applications

    Aniket A Khule, Manoj S Nagmode, and Rajkumar D Komati. Automated object counting for visual inspection applications. In2015 International Conference on Information Processing (ICIP), pages 801–806. IEEE, 2015

    work page 2015

  32. [32]

    Visual question answering: A survey of methods, datasets, evaluation, and challenges.ACM Computing Surveys, 57(10):1–35, 2025

    Byeong Su Kim, Jieun Kim, Deokwoo Lee, and Beakcheol Jang. Visual question answering: A survey of methods, datasets, evaluation, and challenges.ACM Computing Surveys, 57(10):1–35, 2025

    work page 2025

  33. [33]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InEuropean conference on computer vision, pages 740–755. Springer, 2014. 11

    work page 2014

  34. [34]

    Countr: Transformer-based generalised visual counting.arXiv preprint arXiv:2208.13721, 2022

    Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting.arXiv preprint arXiv:2208.13721, 2022

    work page arXiv 2022

  35. [35]

    Countse: Soft exemplar open-set object counting

    Shuai Liu, Peng Zhang, Shiwei Zhang, and Wei Ke. Countse: Soft exemplar open-set object counting. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 21536–21546, 2025

    work page 2025

  36. [36]

    Context-aware crowd counting

    Weizhe Liu, Mathieu Salzmann, and Pascal Fua. Context-aware crowd counting. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5099–5108, 2019

    work page 2019

  37. [37]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoor- thi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99–106, 2021

    work page 2021

  38. [38]

    Omnicount: Multi-label object counting with semantic-geometric priors

    Anindya Mondal, Sauradip Nag, Xiatian Zhu, and Anjan Dutta. Omnicount: Multi-label object counting with semantic-geometric priors. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 19537–19545, 2025

    work page 2025

  39. [39]

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

    T Nathan Mundhenk, Goran Konjevod, Wesam A Sakla, and Kofi Boakye. A large contextual dataset for classification, detection and counting of cars with deep learning. InEuropean conference on computer vision, pages 785–800. Springer, 2016

    work page 2016

  40. [40]

    Can current ai models count what we mean, not what they see? a benchmark and systematic evaluation

    Gia Khanh Nguyen, Yifeng Huang, and Minh Hoai. Can current ai models count what we mean, not what they see? a benchmark and systematic evaluation. In2025 International Conference on Digital Image Computing: Techniques and Applications (DICTA), pages 1–8. IEEE, 2025

    work page 2025

  41. [41]

    Few-shot object counting and detection

    Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. InEuropean Conference on Computer Vision, pages 348–365. Springer, 2022

    work page 2022

  42. [42]

    Situate–synthetic object counting dataset for vlm training

    René Peinl, Vincent Tischler, Patrick Schröder, and Christian Groth. Situate–synthetic object counting dataset for vlm training. In21st International Conference on Computer Vision Theory and Applications, 2026

    work page 2026

  43. [43]

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

    Jer Pelhan, Alan Lukežic, Vitjan Zavrtanik, and Matej Kristan. Dave - a detect-and-verify paradigm for low-shot counting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23293–23302, June 2024

    work page 2024

  44. [44]

    Generalized-scale object counting with gradual query aggregation

    Jer Pelhan, Alan Lukežiˇc, and Matej Kristan. Generalized-scale object counting with gradual query aggregation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 8314–8321, 2026

    work page 2026

  45. [45]

    A novel unified architecture for low-shot counting by detection and segmentation.Advances in Neural Information Processing Systems, 37:66260–66282, 2024

    Jer Pelhan, Alan Lukežiˇc, Vitjan Zavrtanik, and Matej Kristan. A novel unified architecture for low-shot counting by detection and segmentation.Advances in Neural Information Processing Systems, 37:66260–66282, 2024

    work page 2024

  46. [46]

    Cell counting.Current protocols in cytometry, (1):A–3A, 1997

    Mary C Phelan and Gretchen Lawler. Cell counting.Current protocols in cytometry, (1):A–3A, 1997

    work page 1997

  47. [47]

    Iterative crowd counting

    Viresh Ranjan, Hieu Le, and Minh Hoai. Iterative crowd counting. InProceedings of the European conference on computer vision (ECCV), pages 270–285, 2018

    work page 2018

  48. [48]

    Learning to count everything

    Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3394–3403, 2021

    work page 2021

  49. [49]

    Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding

    Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. InProceedings of the IEEE/CVF international conference on computer vision, pages 10912–10922, 2021

    work page 2021

  50. [50]

    Pushing the frontiers of uncon- strained crowd counting: New dataset and benchmark method

    Vishwanath A Sindagi, Rajeev Yasarla, and Vishal M Patel. Pushing the frontiers of uncon- strained crowd counting: New dataset and benchmark method. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 1221–1231, 2019. 12

    work page 2019

  51. [51]

    Jhu-crowd++: Large-scale crowd counting dataset and a benchmark method.IEEE transactions on pattern analysis and machine intelligence, 44(5):2594–2609, 2020

    Vishwanath A Sindagi, Rajeev Yasarla, and Vishal M Patel. Jhu-crowd++: Large-scale crowd counting dataset and a benchmark method.IEEE transactions on pattern analysis and machine intelligence, 44(5):2594–2609, 2020

    work page 2020

  52. [52]

    A low-shot object counting network with iterative prototype adaptation

    Nikola Ðuki ´c, Alan Lukežiˇc, Vitjan Zavrtanik, and Matej Kristan. A low-shot object counting network with iterative prototype adaptation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 18872–18881, 2023

    work page 2023

  53. [53]

    Exploring contextual attribute density in referring expression counting

    Zhicheng Wang, Zhiyu Pan, Zhan Peng, Jian Cheng, Liwen Xiao, Wei Jiang, and Zhiguo Cao. Exploring contextual attribute density in referring expression counting. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 19587–19596, 2025

    work page 2025

  54. [54]

    Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation

    Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 803–814, 2023

    work page 2023

  55. [55]

    Native and Compact Structured Latents for 3D Generation

    Jianfeng Xiang, Xiaoxue Chen, Sicheng Xu, Ruicheng Wang, Zelong Lv, Yu Deng, Hongyuan Zhu, Yue Dong, Hao Zhao, Nicholas Jing Yuan, et al. Native and compact structured latents for 3d generation.arXiv preprint arXiv:2512.14692, 2025

    work page internal anchor Pith review arXiv 2025

  56. [56]

    Structured 3d latents for scalable and versatile 3d generation

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21469–21480, 2025

    work page 2025

  57. [57]

    Weidi Xie, J Alison Noble, and Andrew Zisserman. Microscopy cell counting and detection with fully convolutional regression networks.Computer methods in biomechanics and biomedical engineering: Imaging & Visualization, 6(3):283–292, 2018

    work page 2018

  58. [58]

    Polyhaven: a curated public asset library for visual effects artists and game designers, 2021

    Greg Zaal, Rob Tuytel, Rico Cilliers, James Ray Cock, Andreas Mischok, Sergej Majboroda, Dimitrios Savva, and Jurita Burger. Polyhaven: a curated public asset library for visual effects artists and game designers, 2021

    work page 2021

  59. [59]

    Cross-scene crowd counting via deep convolutional neural networks

    Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang Yang. Cross-scene crowd counting via deep convolutional neural networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 833–841, 2015

    work page 2015

  60. [60]

    Semi- supervised multi-view crowd counting by ranking multi-view fusion models.arXiv preprint arXiv:2512.16243, 2025

    Qi Zhang, Yunfei Gong, Zhidan Xie, Zhizi Wang, Antoni B Chan, and Hui Huang. Semi- supervised multi-view crowd counting by ranking multi-view fusion models.arXiv preprint arXiv:2512.16243, 2025

    work page arXiv 2025

  61. [61]

    Learning to Count Objects in Natural Images for Visual Question Answering

    Yan Zhang, Jonathon Hare, and Adam Prügel-Bennett. Learning to count objects in natural images for visual question answering.arXiv preprint arXiv:1802.05766, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  62. [62]

    Cream green dinosaur figure

    Yuda Zou, Zijian Zhang, and Yongchao Xu. Decoupling what to count and where to see for referring expression counting. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 14113–14121, 2026. 13 A Additional experiments In this section we present and discuss important additional results which were not added to our main paper due...

    work page 2026