NetherlandsForensicInstitute/vuurwerkverkenner-development-data

NFI Fireworks Development Dataset for the "Vuurwerkverkenner" Application

The Netherlands Forensic Institute (NFI) Fireworks development dataset consists of scans of fireworks wrappers from fireworks that were investigated in casework in the Netherlands from 2010 onwards. Artificially created snippets are available for all wrappers, and for a subset of the wrappers photographs of actual fireworks snippets (pieces of the wrapper post-detonation) are included.

Data overview

The data consists of 381 wrappers in 224 categories. There are 40 categories with multiple wrappers, and real snippets are available for 40 wrapper categories.

Data collection

Fireworks wrappers were scanned from all physically available specimens at the NFI or by partners from casework conducted between 2010-2025.

As for the collection of fireworks snippets, fireworks were detonated and the resulting fragments (snippets) gathered manually. The snippets were then cleaned, dried, and categorized by wrapper design. Photographs of the snippets were taken against a white background; each photograph may contain multiple snippet pieces.

Creation of the real snippets was based on the availability of fireworks during the data collection period. To enhance the dataset's diversity and volume, additional wrappers were selected to be printed and affixed to Cobra 6 fireworks, in collaboration with domain experts.

Data structure

The top level of our data structure is organized by the data collection method. For instance, the artificial_snippets folder contains artificially created virtual snippets, whereas real_snippets comprises images of actual exploded fireworks. The mock_snippets folder includes images of real exploded fireworks within more realistic settings, such as with a ruler, varying background colors, or different lighting conditions.

Each data collection method folder is further subdivided into train, test, and validation folders, supporting effective model training. Additionally, the artificial_snippets folder features a reference folder containing data used for calculating reference embeddings.

For the artificial_snippets and real_snippets folders, the third layer groups fireworks by category, based on wrapper similarity. For example, variants from different years, such as Cobra 6 from 2018 and 2019, are grouped together in a category due to similar wrapper design. This classification was established with input from domain experts. The structure differs slightly for the mock_snippets folder, which includes images resembling those from a mock crime scene, and can be directly used as input for the model.

Finally, the lowest layer comprises folders named after specific fireworks wrappers, containing multiple snippet images (e.g., 0.png, 1.png) available for each wrapper.

Below is a visual representation of the data structure:

└───  artificial_snippets (#381 categories)
     └───reference (#35 images per wrapper)
          └───cobrawit
                └───cobra t.t 
                     └───0.png
                     └───1.png
                     └───2.png
          └───butterfly
                └───butterfly 100
                     └───0.png
                     └───1.png
                     └───2.png
     └───train (#35 images per wrapper)
     └───test (#5 images per wrapper)
     └───val (#5 images per wrapper)
└───  mock_snippets (#2489 images)
     └───train (#1741 images)
       └───SH-HNT-VT-M1-2_zwart.jpg
       └───SH-HNT-C8-A1-10.jpg
       └───SH-HNT-C8-A9-11.jpg
     └───test (#374 images)
     └───val (#374 images)
└───  real_snippets (#40 categories)
     └───train(#1-150 images per available category/label)
          └───cobrazwart
                └───super cobra 62g (uk 2014)
                     └───0.png
                     └───1.png
                     └───2.png
          └───jokerroze
                └───joker
                     └───0.png
                     └───1.png
                     └───2.png
     └───test (#1-148 images per available category/label)
     └───val (#4-149 images per available category/label)

Firework wrappers

README.md Images of the scanned wrappers have been resized to a maximum dimension of 2000 pixels in either height or width.

Categorisation

Below is an overview of the number of wrappers found in categories and how many categories contain that amount of wrappers:

Number of wrappers per category:

Training

The methodology adopted by the NFI for model training with this dataset is detailed in the model card of the Vuurwerkverkenner at https://huggingface.co/NetherlandsForensicInstitute/vuurwerkverkenner, also linked to this dataset.

Possible bias

The reference set of firework wrappers reflects the fireworks investigated by the NFI, which may not be representative of those available in other countries or in the future.

The snippet dataset was influenced by the availability of fireworks. Due to limited availability, some snippets result from wrappers glued to Cobra 6 fireworks rather than the original fireworks, which were noted to detach more easily than those on original fireworks, potentially affecting snippet creation.

Very small snippets have not been included in the dataset.