Weapon Visual Detector

General Information

Purpose: To detect the presence and location of a variety of weapons within an image Architecture: Trained using Clarifai's version of InceptionV2 Intended Use: The v1 version of this model is performative in domains such as retail / catalogue imagery, stock photography, or social media type images where the object is one of the main focuses of the model (ie. person posing with a gun) Limitations: This model was not trained on security camera type data (high angle, low resolution, distant/small objects), therefore using it on that domain could have interesting results

Training Data, Evaluation Data & Taxonomy

Training Data

The model is trained on data sourced and labeled by our internal teams leveraging a combination of open sourced datasets, commercially licensed or public domain imagery, as well data from our internal collections. As such, it can not be shared publicly at this time.

Evaluation Data

Eval set 1

• Source: https://www.kaggle.com/ar5p1edy/weapons-datasets
• Covers handgun and long-gun
• Mostly comprised of catalogue or stock photography type imagery. Objects are mainly either isolated and on a solid background, or being prominently displayed as the focus / center of the photograph.
• mAP @ IoU 0.5: 0.89347

Eval set 2

• Source: https://www.kaggle.com/cassdc/people-carrying-ars-handguns-and-unarmed
• Covers handgun and long-gun
• The dataset consisted of two different groups of data: a) people carrying rifles b) people carrying handguns
• Mostly close-up, in-frame images of people “open-carrying” their weapons, generally in a holster or strapped to a person’s back. Images were mostly small/tiny (~300 x 200 pixels) and low resolution (72 dpi). As such, they’re drastically smaller and blurrier than the training images used.
• mAP @ IoU 0.5: 0.68677

Eval set 3

• Source: https://dasci.es/transferencia/open-data/24705/
• Covers knife
• Images are mostly either catalogue types images (object is mostly standalone on a solid background) or being handled by a person, generally from self defense demonstration / education videos.
• mAP @ IoU 0.3: 0.61948 note: metrics are being reported here using an IoU of 0.3, as opposed to 0.5 like the other eval sets, because the ground truth on the standalone knives are only localized to the blades, while our predicted bounding boxes are localized to the entire knife (handle included).

Eval set 4

• Source: hand-sourced by our internal teams
• Covers ammunition, bow/crossbow, handgun, heavy-artillery, long-gun, paintball-gun, sword
• One of the goals here was to create a real-world representative dataset as possible, so there’s little catalogue / stock photography included in here. This eval set should be looked as more of a future goal for later versions of this model.
• mAP @ IoU 0.5: 0.55101

Taxonomy

Format: concept_name [concept_id]

ammunition [ai_8028xs48]
bow/crossbow [ai_hLqzpr3Q]
handgun [ai_LbbcjM2f]
heavy-artillery [ai_HQ02gdLH]
knife [ai_hhtMkWHL]
long-gun [ai_0BF4zq10]
paintball-gun [ai_4qvFwcQJ]
sword [ai_T27fbfxG]