license: eupl-1.2
language:
- nl
NFI Fireworks training dataset for the "Vuurwerkverkenner" application
The Netherlands Forensic Institute (NFI) Fireworks training dataset consists of scans of fireworks wrappers from fireworks that were investigated in casework in the Netherlands between 2010 and 2025. For a subset of these wrappers, photographs of fireworks snippets (pieces of the wrapper after the detonation) were added.
Data overview
The data consists of 347 wrappers in 216 categories. There are 36 categories with multiple wrappers, and snippets are available for 216 wrapper categories.
Data collection
For the fireworks wrappers, all wrappers physically available at the NFI Explosives team from casework between 2010-2025 were scanned.
For the collection of the fireworks snippets, fireworks were exploded and snippets that resulted were gathered by hand. Snippets were cleaned, dried and sorted by wrapper. Snippets were then photographed on a white background. One snippet photo contains multiple snippets.
Sampling was done based on availability of fireworks in the data collection period. To diversify and increase the dataset, additional wrappers were chosen to print and glue to Cobra 6 fireworks. This selection was done in collaboration with domain experts.
Data structure
The root level of our data structure is organized by the method of data collection. For instance, the artificial_snippets folder contains virtual snippets created artificially, while real_snippets comprises images of real exploded fireworks.
Within each data collection method folder, the data is further divided into train, test, and validation folders. These divisions help in training our model effectively. In addition, the artificial_snippets folder includes a reference folder, which holds data used to calculate reference embeddings.
Moving to the third layer, each folder represents a firework category. Firework types are grouped based on the similarity of their wrappers. For example, variants from different years, like Cobra 6 from 2018 and 2019, are considered similar due to their wrapper design. These categorizations were made with the assistance of domain experts.
The final layer contains folders named after specific firework wrappers. Each of these folders includes multiple images of snippets (e.g., 0.png, 1.png) if they are available for that particular wrapper category.
Below is a visual representation of the data structure:
└─── artificial_snippets
└───reference
└───cobrawit
└───cobra t.t.
└───0.png
└───1.png
└───2.png
└───butterfly
└───cobra t.t.
└───0.png
└───1.png
└───2.png
└───supertop
└───train
└───test
└───val
└─── mock_snippets
└───train
└───test
└───val
└─── real_snippets
└───train
└───test
└───val
Firework wrappers
Wrappers have been scanned, and the images have been reduced in size to a maximum of 2000 pixels in either height or width.
Categorisation
Categories with multiple wrappers:
| Category | Number of wrappers |
|---|---|
| atomyc | 4 |
| mascleta | 2 |
| bigboy | 3 |
| dumbumzwart | 12 |
| blackvlinder | 2 |
| blackwidow | 3 |
| bluestorm1 | 2 |
| caramella2 | 3 |
| chooet | 2 |
| cobrazwart | 63 |
| zink2 | 3 |
| zink3 | 3 |
| tuonogolf | 4 |
| tuonorood | 2 |
| tuonovuur | 2 |
| tuonogoud | 2 |
| viperzwart | 4 |
| xplode | 2 |
| cobrazilver | 5 |
| crazyrobots | 2 |
| delovarana1 | 2 |
| delovarana2 | 2 |
| thunderstorm | 2 |
| dumbumzwartrood | 2 |
| dumbumzwartzilver | 3 |
| jorge2 | 2 |
| horror | 2 |
| jr101 | 2 |
| lupo | 3 |
| match_cracker1 | 2 |
| retorno2 | 2 |
| petarde | 2 |
| profi | 2 |
| shark2 | 2 |
| scream | 10 |
| shark1 | 2 |
Categories with snippet collections:
| Category | Number of snippet collections |
|---|---|
| match_cracker2 | 33 |
| shell2 | 33 |
| shell1 | 33 |
| 3040 | 33 |
| ghost1 | 33 |
| achtung2 | 33 |
| achtung_skull | 33 |
| achtung1 | 33 |
| alarm5 | 33 |
| appletrate | 33 |
| aquilacentinaio | 33 |
| zorro | 33 |
| atomyc | 132 |
| mascleta | 66 |
| bigboy | 99 |
| dumbumzwart | 396 |
| bigjumbo | 33 |
| bigthunder | 33 |
| bigtigertropic | 33 |
| blackdeath | 33 |
| blackthunder1 | 33 |
| blackthunder2 | 33 |
| blackthunder3 | 33 |
| blackvlinder | 66 |
| blackwidow | 99 |
| blitzschlag | 33 |
| bluestorm1 | 66 |
| bluestorm2 | 33 |
| zink1 | 33 |
| bomb1 | 33 |
| bomb2 | 33 |
| butterfly | 33 |
| caramella1 | 33 |
| caramella2 | 99 |
| celebrationcracker | 33 |
| chooet | 66 |
| circoblitz | 33 |
| cobrazwart | 2079 |
| zink2 | 99 |
| zink4 | 33 |
| zink3 | 99 |
| zink5 | 33 |
| signalrakete901blauw | 33 |
| silverdemon | 33 |
| siouxwarior | 33 |
| sp1010 | 33 |
| spanishcracker | 33 |
| stielhandgranate | 33 |
| succubus | 33 |
| supercracker | 33 |
| superexplosion2 | 33 |
| superexplosion4 | 33 |
| superexplosion1 | 33 |
| superexplosion3 | 33 |
| supertop | 33 |
| thumpingthunder | 33 |
| thundercracker | 33 |
| thunderking | 33 |
| tigerboom | 33 |
| titan | 33 |
| shell5 | 33 |
| tp2 | 33 |
| tuonoblauw | 33 |
| tuonogolf | 132 |
| tuonorood | 66 |
| tuonovuur | 66 |
| tuonogoud | 66 |
| tuonott4 | 33 |
| txp001 | 33 |
| viperbruin | 33 |
| vipergeel | 33 |
| viperzwart | 132 |
| viperrood | 33 |
| viper2 | 33 |
| virus3 | 33 |
| virus | 33 |
| vlinder1 | 33 |
| xplode | 66 |
| yellowshock | 33 |
| bulldog | 33 |
| mario200 | 33 |
| monster1 | 33 |
| viperduo | 33 |
| virus2 | 33 |
| blackthunder4 | 33 |
| euphoria | 33 |
| bigboyboom | 33 |
| duxmealux | 33 |
| fotzenblitz | 33 |
| padrino | 33 |
| numero40 | 33 |
| numberone2 | 33 |
| shell4 | 33 |
| gigantmaroon2 | 33 |
| gun | 33 |
| redbutterfly | 33 |
| gretathunder1 | 33 |
| gretathunder2 | 33 |
| vitaminf50 | 33 |
| tirex | 33 |
| panzerfaust1 | 33 |
| cobrazilver | 165 |
| cobrawit | 33 |
| cobratraat1 | 33 |
| cobratraat2 | 33 |
| colorreport | 33 |
| colourflower | 33 |
| bp0038p | 33 |
| coloursalute | 33 |
| corona | 33 |
| torpedo | 33 |
| crazybang | 33 |
| crazyrobots | 66 |
| delovarana3 | 33 |
| delovarana1 | 66 |
| delovarana2 | 66 |
| diablo (blauw) | 33 |
| diablo (rood) | 33 |
| diablo (zwart) | 33 |
| dieptebom | 33 |
| dinamite | 33 |
| thunderstorm | 66 |
| doppelschlag | 33 |
| dragonboom1 | 33 |
| dragonboom2 | 33 |
| dumbumvuurretorno100 | 33 |
| dumbumzwartrood | 66 |
| dumbumoranje | 33 |
| dumbumzwartzilver | 99 |
| echobomb | 33 |
| ekstra_cobra | 33 |
| cometa | 33 |
| matabrujas | 33 |
| raptor | 33 |
| flashing thunder | 33 |
| folgore | 33 |
| nitraat9 | 33 |
| jorge1 | 33 |
| jorge2 | 66 |
| funkenschlag1 | 33 |
| funkenschlag2 | 33 |
| ghost2 | 33 |
| gigantmaroon1 | 33 |
| gladiator1 | 33 |
| gladiator2 | 33 |
| goldthunder | 33 |
| gorillabomb | 33 |
| handgranade | 33 |
| horror | 66 |
| joker2 | 33 |
| joker4 | 33 |
| jokerroze | 33 |
| joker1 | 33 |
| joker3 | 33 |
| jr101 | 66 |
| kalasnikov | 33 |
| kittywhr | 33 |
| nitraat5 | 33 |
| 817 | 33 |
| krachmen | 33 |
| labomba | 33 |
| littlecooper | 33 |
| lupo | 99 |
| match_cracker1 | 66 |
| megabomb | 33 |
| retorno2 | 66 |
| batterij fb | 33 |
| nitraat11 | 33 |
| nitraat6 | 33 |
| napolitaansebom | 33 |
| newrambo | 33 |
| nitraat2 | 33 |
| nitraat3 | 33 |
| nitraat4 | 33 |
| nitrobomb | 33 |
| numberone1 | 33 |
| cs1 | 33 |
| p1000 | 33 |
| nitraat8 | 33 |
| bombbon | 33 |
| 859 | 33 |
| nitraat7 | 33 |
| petarde | 66 |
| nitraat1 | 33 |
| profi | 66 |
| shark2 | 66 |
| pyropower | 33 |
| pyrostar | 33 |
| rambo | 33 |
| rastatraat | 33 |
| reaper | 33 |
| redbaron | 33 |
| reddevil | 33 |
| redflower | 33 |
| redthunder | 33 |
| retorno3 | 33 |
| retorno1 | 33 |
| rex1 | 33 |
| rocket2 | 33 |
| rocket3 | 33 |
| saperp2000 | 33 |
| schockerxxl | 33 |
| scream | 330 |
| screamineagle | 33 |
| shark1 | 66 |
| shell3 | 33 |
| barracuda | 33 |
| purplestorm | 33 |
| pitbull | 33 |
| lupoblauw | 33 |
| pertardof2 | 33 |
| nebelschlagpaars | 33 |
| nebelschlagrood | 33 |
| nebelschlaggeel | 33 |
| nebelschlagblauw | 33 |
| explod | 33 |
Training
The steps followed at the NFI to actually train the model with these data, are described in the model card of the Vuurwerkverkenner at https://huggingface.co/NetherlandsForensicInstitute/vuurwerkverkenner (also linked to this data set).
Possible bias
The reference set of firework wrappers is dependent on which fireworks were investigated by the NFI. This may not be representative for fireworks in other countries or in the future.
Data collection of snippets was influenced by the availability of fireworks. Due to low availability, some snippets are the result of wrappers glued to Cobra 6 fireworks instead of the original fireworks. We noticed that these wrappers tended to detach more easily than wrappers on original firework. This may have influenced the snippet creation.
Very small snippets were not included in our dataset.