Scylla AI Evidence: Benchmarks, Testing & Case Studies
Reviewed on October 15, 2026.
Scylla develops AI video analytics for physical security, including threat detection, object tracking and video investigation. Scylla supports edge, on-premises, hybrid and air-gapped deployments, allowing organizations to align AI video analytics with their security, connectivity and data sovereignty requirements. This page brings together public benchmark results, government reporting, customer case studies and technical testing to help organizations evaluate the evidence behind Scylla’s capabilities.
Each result applies to the task, dataset or deployment described. Public benchmarks, operational evaluations and internal tests answer different questions and should be assessed in their original context.
Public computer vision benchmarks
ScyllaNet: first-place COCO object detection result
Precision across IoU thresholds from 0.50 to 0.95. The published result also reports first place in nine of the twelve evaluation metrics.
COCO evaluates how accurately a model detects and locates objects in images. This result provides evidence of object-detection performance on a standardized benchmark. It is not a measurement of firearm-detection accuracy or end-to-end alert latency.
MOT17 and MOT20: multi-object tracking achievements
Scylla’s published tracking research reports first-place results in the MOT17 and MOT20 competitions hosted on CodaBench. Its MOT17 summary reports an IDF1 score of 87.35 and 708 identity switches for SCYLLA-MOT.
These benchmarks assess how consistently a system follows people across video frames, including through occlusion and crowded scenes. Tracking identity means maintaining a consistent track for a person; it does not require identifying that person by name.
Read the multi-object tracking research and benchmark references
Rankings describe the published achievements and may change as new submissions are evaluated. Refer to the linked results for benchmark scope and submission details.
Published research: SIoU loss
Zhora Gevorgyan’s research paper, “SIoU Loss: More Powerful Learning for Bounding Box Regression,” was submitted to arXiv on May 25, 2022. It introduces a bounding-box regression loss that incorporates the direction of the mismatch between predicted and target boxes, alongside other localization factors.
Government-reported operational evaluation
Scylla at Blue Grass Army Depot
The U.S. Army documented testing and a demonstration of Scylla AI at Blue Grass Army Depot in 2024. Its October 8 report described alerts within seconds for two individuals carrying weapons, a perimeter-breach scenario, smoke detection and a simulated fight.
A separate Defense Department report quoted the depot’s Electronic Security Services manager as reporting a probability of detection above the 96% standard during Scylla’s test and evaluation, together with a reduction in false alarms caused by environmental conditions.
The reported findings apply to that evaluation and its scenarios; they do not establish a universal accuracy rate for every Scylla module or deployment.
Read the U.S. Army report: BGAD tests AI for enhanced security
Read the Defense Department report on AI physical-security testing
Customer case studies: reducing false alarms
nuisance alarms and help operators focus on events requiring review.
| Customer | Location | Reported false-alarm reduction |
|---|---|---|
| Sentinel Monitoring Ltd. | United Kingdom | Up to 99% |
| Customer | Location | Reported false-alarm reduction |
|---|---|---|
| Raid Security | South Africa | Up to 98% |
| Totem | Ecuador | Up to 97% |
Sentinel Monitoring Ltd.
The published case study reports filtering of up to 99% of false alarms, including nuisance activations associated with foliage, changing shadows and animals.
Read the Sentinel Monitoring case study
Raid Security
Raid Security adopted Scylla to address high alarm volumes and control-room operator fatigue as its camera estate expanded. The published case study reports false-alarm reduction of up to 98% across multiple instances.
Read the Raid Security case study
Totem
Totem implemented Scylla False Alarm Filtering to support its monitoring operations in Ecuador. The published case study reports filtering of up to 97% of false alarms.
These are deployment-specific results published by Scylla. False-alarm reduction measures the change in nuisance alerts; it is not equivalent to detection accuracy, and results vary with site conditions and configuration.
Technical whitepaper: module-level testing
The Scylla technical whitepaper describes internal evaluations, camera requirements and operating conditions for individual modules. The following figures are reported test-set results, rather than independent certification or guaranteed performance at every site.
| Module | Precision | Recall | F1 score |
|---|---|---|---|
| Fight detection | 94.4% | 94.7% | 94.6% |
| Slip and fall detection | 98.8% | 98.2% | 98.5% |
| Suspicious shopping behavior | 97.8% | 88.3% | 92.8% |
| Smoke and fire detection | 99.1% | 99.0% | 99.05% |
Precision measures how many positive predictions are correct. Recall measures how many labeled target examples are detected. F1 summarizes the balance between precision and recall.
These results come from separate module test sets and should not be used to rank the modules against one another. For deployment assessment, evaluate missed events and false alarms alongside detection speed and camera suitability.
Visible firearm detection under defined conditions
Scylla’s whitepaper evaluates visible firearm detection across different weapon types, lighting conditions, distances and pointing directions. Tests include daylight and infrared scenarios, with detection assessed over specified time windows.
In the well-lit, light-background test, the reported handgun detection sensitivity within two seconds was 98% at four meters, 95% at eight meters and 92% at twelve meters. These figures describe the specified internal test conditions. Performance depends on weapon visibility, image detail, viewing angle and lighting.
How to validate Scylla for your environment
A site-specific proof of concept should assess the intended cameras, use cases and response workflow under representative conditions. Scylla’s technical guidance supports evaluation of:
● Correct detections and missed events, using labeled target scenarios. ● False alarms per camera per day, including ordinary activity that may resemble a target event. ● Time from a visible event to an alert reaching the operator. ● Camera resolution, object size, illumination, viewing angle and occlusion. ● Model version, thresholds, hardware and integration settings. ● For EDI Assist, also record the model, prompt, prompt type and analysis interval. EDI Assist evaluates individual images, so a brief event may occur between checks. When evaluating EDI Filter, measure both false alarms removed and genuine alarms retained.
Frequently asked questions
What evidence supports Scylla AI’s performance?
Scylla’s evidence includes published COCO, MOT17 and MOT20 benchmark achievements, U.S. government reporting on the Blue Grass Army Depot evaluation, named customer case studies and internal module tests in its technical whitepaper. Each source supports a specific capability or operating context.
Has Scylla AI been tested by the U.S. Army?
Yes. The U.S. Army reported testing and a demonstration of Scylla at Blue Grass Army Depot in 2024. The documented scenarios included armed individuals, a perimeter breach, smoke and a simulated fight. These reports concern the depot’s evaluation rather than a blanket endorsement of all Scylla products.
How much can Scylla reduce false alarms?
Published Scylla case studies report reductions of up to 99% for Sentinel Monitoring, up to 98% for Raid Security and up to 97% for Totem. These are results from particular deployments. A proof of concept should measure false alarm reduction together with the retention of genuine alerts.
Do COCO and MOT results measure weapon-detection accuracy?
No. COCO evaluates general object detection, while MOT17 and MOT20 evaluate multi-object tracking. Weapon-detection performance requires separate testing of visible firearms under relevant camera, lighting and scene conditions.
Does Scylla have one accuracy score for the entire platform?
No. Scylla modules perform different tasks and are evaluated using different datasets and metrics. Detection precision, recall, tracking identity consistency, false-alarm reduction and alert latency describe different aspects of performance. Results should be interpreted for the relevant module and test conditions.
Can Scylla detect concealed weapons or predict an incident?
Scylla’s video analytics analyze visible content in camera footage. Visual weapon detection requires a weapon to be visible and sufficiently distinguishable. It does not see through clothing or walls, or establish what will happen next.
Evaluate Scylla against your operational requirements
Discuss your priority use cases, camera coverage, deployment architecture and evidence requirements with Scylla. A technical consultation can help define a proof of concept with measurable acceptance criteria for your environment.
