Multi-Domain Targeting Alignment

Problem Statement

“In support of the transition to a hybrid fleet and enhanced target classification across all domains, identify Automatic Target Recognition (ATR) technologies capable of detecting, classifying, and distinguishing friendly, adversarial, and high-interest targets across multiple platforms (manned and unmanned) and sensor domains.”

Traditional automatic target recognition models struggle to accurately detect, classify and distinguish targets in complex, multi-domain environments.

This is a personal pain point. During my master's thesis at Cornell Tech's AI Lab, our team developed a solution that used the YOLO algorithm to track target objects, and noticed the accuracy limitations of existing AI models in multi-domain situations.

Limitations with the Current Approach

Across multi-domains such as land, air and sea, most published datasets are narrow in scope (limited object types, resolution, or scale), resulting in limited model accuracy and generalizability.

Dataset	Model	Images / Classes	Metric	Result
Multi-Scale Marine SSS	YOLO	1 584 / 5	mAP_0.5	92.4 %
Trash_ICRA19	YOLOv5-Improved	6 540 / 3	mAP	90.6 %
RUOD	YOLOF	14 000 / 10	mAP_0.5:0.95	50.1 %

The biggest bottleneck to improving model accuracy is the data. In order to reduce classification errors and accelerate information sharing, there is a need to curate relevant training and evaluation datasets for broad anomaly coverage.

Downsides of Manual Data Curation And Labeling

In order to curate training datasets and evaluation datasets for multi-domain tracking, the DOD currently has people manually going through data to annotate images and other unstructured data types with categories and bounding boxes. This has the following downsides:

Issue	Description
Time Consuming	Slow, even with an unlimited labor force, it can take months to years
Costly	Expensive, takes up a significant portion of a team’s budget
Tedious	Laborious, mundane and boring work that people shouldn’t have to do
Requires Expertise	Hard, requires highly trained subject matter expertise with domain-specific experience to find information
Iterative	Repetitive, when new data or new categories are added, or business requirements change, humans need to manually go through the data

Business Model and Value Add

A really accurate AI solution can help automate the tagging of this information and would reduce operator burden and increase autonomy. By leveraging Anote’s AI solution, the Navy can reduce data processing time from months to weeks, mitigate the risk of human error, and rapidly adapt to new categories or requirements to improve operational resilience and model accuracy.

Metric	Current Process	Anote Solution
Personnel	1,200	50
Annual Cost / Person	$80,000	$80,000
Product / PoC Cost	—	$X
Time to Label	1 yr	1 mo
Annual Spend	$96 M	$4.2 M

Market Opportunity and Dual Use Potential

Our team has done a lot of research on benchmark datasets across land, air and sea. Below you can see a few dual use cases that our team has been benchmarking to show how these models can operate in complex, real-world conditions.

Civil / Federal Domain	Representative Use-Case	Key ATR Requirement
Domestic Service Robot	Retrieve laundry; load dishes in cluttered homes	Robust close-range detection under occlusion & variable lighting
Army AI Grand Challenge	Sub-second hostile-asset detection in aerial & ground feeds	Low-latency inference; camouflage resilience
NGA Satellite Pipeline	National-scale mapping & change detection	Wide-area scanning; mis-label audit
NYC Subway Security	Real-time threat & crowd-flow monitoring	Weapon recognition in crowded scenes; anomaly detection