This model identifies pistols and long guns in images. It accepts any jpg or png images as input, although it works best with at least 480p resolution. It outputs a JSON containing the location, confidence, and type of firearm (pistol or long gun). The model works even in dark, blurry, and occluded images, and is effective in both real-time camera monitoring and forensic data analysis.
Many models are available for limited use in the free Modzy Basic account.
This model was trained and tested using Actuate’s proprietary in-house dataset of 500k+ gun images. This model used 50,000 unique images for training, and another 100,000 for testing, including over 1,000 security camera videos. The model achieves 98% Precision and 96% Recall on the full test dataset on a per-image level. When images are treated as components of videos and results are combined, the model achieves 99%+ Recall identifying guns in test videos. This model’s strengths include its capability to detect pistols and long guns in difficult environments, including low resolution images, night vision scenarios, poor weather, at long ranges, and when firearms are mostly occluded. This model works on both CPUs and GPUs, but GPUs are recommended for real-time analysis.
99% Accuracy – The fraction of correct predictions made by the classifier. This metric is calculated by dividing the number of correct predictions by the total number of predictions.
97% F1 Score -Is the harmonic mean of the precision and recall, with best value of 1. It measures the balance between the two metrics.
This model utilizes Actuate’s proprietary fork of the YOLOv4 one-stage detection model, which is an open-source machine learning platform. Actuate’s customizations include modifications to the model’s underlying Darknet framework, network structure, scaling, and pooling parameters. This allows the model to detect firearms in complex, poor quality images while maintaining real-time performance. This version of the model runs using PyTorch to maintain compatibility across systems.
This model was trained and tested using Actuate’s proprietary in-house dataset of 500k+ gun images. This model used 50,000 unique images for training after applying transfer learning from COCO, an open-source dataset containing over 1.5M object instances. The model obtains near-perfect 99%+ Recall for detecting both pistols and long guns when images are treated as components of videos, with precision over 98%. Training took 50 hours on 8x Nvidia Tesla V100 GPUs. The model was trained using stochastic gradient descent with dynamic learning rates.
Get a video demo and join the community of developers and customers building the future of Artificial Intelligence.