ScrewCount

ScrewCount is a dataset for dense small-object counting in industrial inspection settings, designed to benchmark both exemplar-based few-shot counting and text-guided object counting. It focuses on challenging manufacturing scenarios with small, overlapping, densely packed, and visually similar objects, specifically screws and nuts.

The dataset was introduced in the paper: ScrewCount: A Dataset and Benchmark for Exemplar Efficiency and Text-Guided Few-Shot Object Counting

Supported Tasks and Leaderboards

This dataset can be used for:

• Few-shot object counting
• Exemplar-based counting
• Text-guided counting
• Dense small-object counting
• Point-supervised counting
• Density map regression
• Industrial object detection (limited, depending on annotation usage)

Evaluation Metrics

The benchmark primarily uses:

• MAE (Mean Absolute Error)
• RMSE (Root Mean Squared Error)

These are standard metrics for object counting tasks.

Dataset Structure

Each sample consists of a high-resolution image of densely packed industrial objects, along with annotations that can support multiple counting paradigms.

A typical sample may include:

• image: RGB image
• points: point annotations for object instances
• density_map: generated density map corresponding to point annotations
• exemplar_boxes: up to 6 exemplar bounding boxes per image for few-shot counting
• split: train / val

Dataset Summary

Curation Rationale

Most existing counting datasets focus on crowds, cells, or natural scenes. These benchmarks do not fully capture the challenges of industrial inspection, where objects are often:

• very small,
• densely packed,
• heavily overlapping,
• partially occluded,
• and visually similar.

ScrewCount was created to address this gap and provide a benchmark tailored to real-world industrial counting scenarios where annotation is expensive and few-shot methods are especially relevant.

Source Data

The dataset contains high-resolution 3024×4032 images of:

• screws
• nuts

The objects vary in:

• size: small, medium, large
• shape
• color: white and black

Images represent dense manufacturing-like scenes with object counts often in the hundreds per image.

Categories

Category	Training Samples	Test Samples
Screw	500	100
Nuts	500	100

Summary

• 2 object categories
• 1000 training images total
• 200 test images total
• Image resolution: 3024×4032
• Objects per image: typically a few hundred
• Exemplars per image: 6

Contributions

Contributions, issue reports, and benchmark reproductions are welcome through the associated repository.

Citation

If you use this dataset, please cite:

@conference{visapp26,
author={Farnaz Delirie and Afshin Dini and Amirmasoud Molaei and Leila Sadeghi},
title={ScrewCount: A Dataset and Benchmark for Exemplar Efficiency and Text-Guided Few-Shot Object Counting},
booktitle={Proceedings of the 21st International Conference on Computer Vision Theory and Applications - Volume 1: VISAPP},
year={2026},
pages={306-313},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0014236300004084},
isbn={978-989-758-804-4},
issn={2184-4321},
}

License

This model is licensed under the Attribution–NonCommercial 4.0 International License (CC BY-NC 4.0).