README.md

---
license: cc-by-4.0
dataset_info:
  features:
    - name: image
      dtype: image
    - name: file_name
      dtype: string
    - name: total_count
      dtype: int64
    - name: num_classes
      dtype: int64
    - name: class_names
      sequence: string
    - name: class_counts
      sequence: int64
    - name: class_descriptions
      sequence: string
    - name: objects
      struct:
        - name: bbox
          sequence:
            sequence: float64
        - name: category
          sequence: int64
configs:
  - config_name: default
    data_files:
      - split: train
        path: parquet_data/train-*
      - split: validation
        path: parquet_data/val-*
      - split: test
        path: parquet_data/test-*
tags:
  - counting
  - synthetic
  - computer-vision
  - open-vocabulary
  - segmentation
pretty_name: MixCount
task_categories:
  - image-classification
  - object-detection
---


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

<p align="center">
  <a href="https://corentindumery.github.io/">Corentin Dumery*</a> · <a href="https://reuben.ox.ac.uk/people/niki-amini-naieni">Niki Amini-Naieni*</a> · <a href="https://www.linkedin.com/in/shervin-n">Shervin Naini</a> · <a href="https://people.epfl.ch/pascal.fua">Pascal Fua</a>
  <br>
  <sub>EPFL · University of Oxford · Northwestern University · (* equal contribution)</sub>
</p>

<p align="center">
  <a href="https://arxiv.org/abs/2605.18063"><img src="https://img.shields.io/badge/Paper-arXiv-b31b1b?style=for-the-badge" alt="Paper"></a>
  <a href="https://corentindumery.github.io/projects/mixcount.html"><img src="https://img.shields.io/badge/Project-Page-143038?style=for-the-badge" alt="Project page"></a>
</p>

<style>
.mixcount-stats {
  display: grid;
  grid-template-columns: repeat(5, minmax(0, 1fr));
  gap: 10px;
  max-width: 900px;
  margin: 0 auto 1.5em;
}
.mixcount-stat {
  background: linear-gradient(180deg, #f8fbfc 0%, #eef4f7 100%);
  border: 1px solid #d8e4ea;
  border-radius: 10px;
  padding: 0.85em 0.5em;
  text-align: center;
}
.mixcount-stat b {
  display: block;
  font-size: 1.35em;
  color: #143038;
}
.mixcount-stat span {
  display: block;
  margin-top: 0.3em;
  font-size: 0.82em;
  color: #4f6468;
  line-height: 1.3;
}
.mixcount-lead {
  max-width: 820px;
  margin: 0 auto 1.5em;
  padding: 1em 1.15em;
  background: #f7fafb;
  border-left: 4px solid #5ba4c4;
  border-radius: 0 8px 8px 0;
  line-height: 1.55;
}
.mixcount-compare-wrap {
  display: flex;
  justify-content: center;
  margin: 0 auto 1.25em;
  overflow-x: auto;
}
.mixcount-compare-wrap table {
  margin: 0 auto;
}
@media (max-width: 720px) {
  .mixcount-stats { grid-template-columns: repeat(2, 1fr); }
}
</style>

<p align="center">
  <img src="images/teaser.png" width="96%" alt="MixCount sample scenes (2×6 grid)">
</p>

<div class="mixcount-lead">
  <b>MixCount</b> is a large-scale synthetic dataset for <b>mixed-object, open-vocabulary counting</b>, the setting that dominates industrial inspection and sorting, but breaks current counting models. Our automatic generation pipeline produces pixel-perfect labels, text prompts at several levels of detail, and visual exemplars at scale.
</div>

<div class="mixcount-stats">
  <div class="mixcount-stat"><b>58K</b><span>counting scenes</span></div>
  <div class="mixcount-stat"><b>1,522</b><span>object classes</span></div>
  <div class="mixcount-stat"><b>4M+</b><span>counting instances</span></div>
  <div class="mixcount-stat"><b>−18.3%</b><span>MAE on PairTally (train)</span></div>
  <div class="mixcount-stat"><b>−20.14%</b><span>MAE on FSC-147 (train)</span></div>
</div>

## Usage

```python
import matplotlib.pyplot as plt
import matplotlib.patches as patches
from datasets import load_dataset

dataset = load_dataset("CorentinDumery/MixCount", split="train", streaming=True)
example = next(iter(dataset))

for name, count, desc in zip(example['class_names'], example['class_counts'], example['class_descriptions']):
    print(f" - {name}: {count} instance(s)")
    print(f"   Description: {desc}")

objects = example['objects']

fig, ax = plt.subplots(1, figsize=(10, 8))
ax.imshow(example['image'])

for bbox, category_id in zip(objects['bbox'], objects['category']):
    x, y, w, h = bbox
    color = plt.colormaps['tab20'](category_id % 20)
    rect = patches.Rectangle(
        (x, y), w, h,
        linewidth=2,
        edgecolor=color,
        facecolor='none',
    )
    ax.add_patch(rect)

plt.axis('off')
plt.show()
```

## Overview

Object counting models often struggle in **mixed-object scenes**. Common failure modes include:

- **(a)** Distinguishing **visually similar objects** (e.g. *big marbles* in PairTally)
- **(b)** Recognizing **self-similar components** as a single entity (e.g. counting pairs of sunglasses rather than lenses)
- **(c)** Ignoring **repetitive background patterns** and focusing on the queried object class

MixCount combines the scale of synthetic datasets with the photorealism of real-world 3D captures while targeting these failure modes. Training on MixCount yields about **20% lower error** on recent open-vocabulary counting benchmarks.

<p align="center">
  <img src="images/bridging.png" width="68%" alt="Training on MixCount improves CountGD++ on PairTally, FSC-147, and MixCount">
</p>

## Dataset overview

<div class="mixcount-compare-wrap">

| | FSC-147 | PairTally | MCAC | **MixCount** |
|---|:---:|:---:|:---:|:---:|
| Multiple object types per image | | ✓ | ✓ | **✓** |
| Fine-grained text prompts | | ✓ | | **✓** |
| External visual exemplars | | | | **✓** |
| Segmentation & bounding boxes | | | ✓ | **✓** |
| # images | 6,135 | 681 | 20K | **58,000** |
| # object classes | 147 | 98 | 343 | **1,522** |

</div>


**Visual & text inputs.** Multiple visual exemplars per object (external crops and in-scene crops at different scales), together with **short, concise, and detailed** text descriptions for flexible open-vocabulary counting prompts.

<p align="center">
  <img src="images/features.png" width="65%" alt="MixCount exemplars and tiered text descriptions">
</p>

**Dense annotations.** Pixel-perfect counting supervision plus instance and class segmentations, bounding boxes, depth, and normal maps.

<p align="center">
  <img src="images/annotations.png" width="65%" alt="MixCount dense annotations">
</p>

**Automatic generator.** Objects, distractors, environment, and camera placement are sampled procedurally to create photorealistic training scenes from high-quality real-world captures of objects, materials, and lighting.

<p align="center">
  <img src="images/data_generator.png" width="65%" alt="MixCount data generation pipeline">
</p>


See the [project page](https://corentindumery.github.io/projects/mixcount.html) and [paper](https://arxiv.org/abs/2605.18063) for additional details.

## Citation

```bibtex
@article{dumery2026mixcount,
  title   = {The MixCount Dataset: Bridging the Data Gap for Open-Vocabulary Object Counting},
  author  = {Dumery, Corentin and Amini-Naieni, Niki and Naini, Shervin and Fua, Pascal},
  journal = {arXiv preprint arXiv:2605.18063},
  year    = {2026}
}
```

## Acknowledgements

We thank [DTC](https://www.projectaria.com/datasets/dtc/), [VasTextures](https://sites.google.com/view/infinitexture/home), [LavalIndoor](http://hdrdb.com/indoor/), and [PolyHaven](https://polyhaven.com/), as well as the [Blender Foundation](https://www.blender.org/). We also thank Andrew Zisserman for insightful discussions. This work is partially funded by the Swiss National Science Foundation, an AWS Studentship, the Reuben Foundation, a Qualcomm Innovation Fellowship (mentors: Dr Farhad Zanjani and Dr Davide Abati), and the AIMS CDT program at the University of Oxford.