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

Corentin Dumery* · Niki Amini-Naieni* · Shervin Naini · Pascal Fua
_{EPFL · University of Oxford · Northwestern University · (* equal contribution)}

MixCount is a large-scale synthetic dataset for mixed-object, open-vocabulary counting, 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.

58Kcounting scenes

1,522object classes

4M+counting instances

−18.3%MAE on PairTally (train)

−20.14%MAE on FSC-147 (train)

Usage

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.

Dataset overview

	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

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.

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

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.

See the project page and paper for additional details.

Citation

@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, VasTextures, LavalIndoor, and PolyHaven, as well as the Blender Foundation. 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.