Publishing Disclaimer: In all of its publications and products, Military Review presents professional information. However, the views expressed therein are those of the authors and are not necessarily those of the Army University, the Department of the Army, or any other agency of the U.S. government.

---

 

Adaptive AI Supremacy Through a Resilient Distributed Model Repository

 

Lt. Col. Eric Sturzinger, PhD, US Army
Capt. William Cocke, PhD, US Army

 

Download the PDF

 

The US Army will soon experience an artificial intelligence (AI) model management problem that promises to escalate as soldiers wield an ever-increasing number of AI-powered tools. Tactical units will rely upon a robust set of heterogeneous AI-enabled systems to maintain decision dominance, for example, through augmented reality goggles, counter-unmanned aircraft system smart scopes, reconnaissance drones, or large language models.¹ These platforms leverage diverse sensor packages with various input data modalities, such as audio, visual, or radio frequency. In addition, due to unique data-sensing perspectives, including satellite, aerial drone, ground vehicle, or stationary asset, each model needs to be trained, deployed, and updated independently. Effectively managing the diverse set of AI models necessitates the employment of a distributed and replicated AI model repository, an Army version of the popular public model repository Hugging Face.² A repository is a software system that manages version control and provides centralized access to users. With the anticipated deployment of many heterogeneous AI models, the need to maintain a unique local model repository is exacerbated by the expectation that peer adversaries will engage in various tactics to disrupt detection by AI systems. The Army is already preparing for the AI/counter-AI cycle across combat domains.³ It must therefore provide units the tools to adapt to such enemy tactics, which includes the ability to fine-tune models to respond to enemy actions. This allows for an AI-enabled Army to rapidly adapt to a dynamic, complex, and unpredictable operational environment.

Cloud-hosted centralized repositories are the preferred industry solution for storing and accessing models. While the cloud affords high availability, standardization, and virtually infinite resources, it suffers from the lack of physical and logical proximity to operations. The Army cannot exclusively rely on ubiquitous cloud connectivity to access model repositories under expected degraded, denied, intermittent, and low-bandwidth (DDIL) conditions.⁴ Rather, units must host a local repository on their tactical server infrastructure, which is replicated with the cloud instance when connectivity permits, to maintain the ability to perform local AI operations. A local repository allows units to share (transfer) their newly derived local knowledge from novel enemy threats detected in their area of operations (AO) in the form of model updates with sister units and higher headquarters for aggregation and distribution. The Army’s distributed model repository will allow knowledge learned by one unit’s experience to propagate throughout the force to ensure that every AI-enabled system has access to the most current knowledge on enemy appearance, behavior, and other tactics, techniques, and procedures (TTP).

A distributed and hierarchically replicated DDIL-resilient AI model repository is critical to the Army’s future ability to conduct AI-enabled operations at scale. Without a robust architecture such as the one described herein, AI/machine-learning operations will be dependent on the cloud, model updates may only come once a year from home station (under the assumption of requisite data availability), and model parameters will not be shared efficiently and rapidly between units. This does not facilitate an Army that transforms in contact—it reinforces the status quo. The Army must employ a distributed, DDIL-resilient AI model repository framework that enables units to share derived knowledge from local experience to maintain adaptive AI supremacy and tactical agility in the intelligence-saturated battlefields of the future.

Heterogeneous AI Demands and Adapting AI to Overcome Enemy TTPs

From reconnaissance to fires to staff processes and predictive logistics, a truly AI-enabled Army will be inundated with AI tools that power intelligent and autonomous systems, ultimately providing commanders decision dominance. Any platform that senses or maneuvers and any intellectual task that analyzes information or produces knowledge is likely to leverage some type of AI model, from incredibly simple identification tasks to complex multiagent systems that interface with user inputs. Even when limited to a single modality, such as visible spectrum cameras, AI models are highly dependent on context. Independently trained models are thus required for systems that process visible data from drones, ground vehicles, and static platforms. The Army already leverages radio frequency, radar, infrared, and other sensor modalities to provide soldiers with real-time situational awareness; each of these is associated with multiple AI models. The sheer number of diverse tasks performed by AI-enabled systems, relying on heterogeneous sensor input modalities in various contexts, will cause the required set of unique AI models employed by units to only increase with time.

The demand for a robust distributed model repository is driven by two fundamental characteristics of the future operational environment:

1. Battlefield ubiquity of heterogeneous AI-enabled systems powered by diverse models and data across various modalities and perspectives.
2. Expected enemy changes to appearance, behavior, and other TTPs requiring the modification of model parameters to prevent unacceptable degradation in AI system performance.

Consider the recently announced antidrone response team.⁵ Soldiers on the team leverage one AI model for infrared sensing and another model to detect electronic threats. In addition, the interceptor drone features an autonomous flight mode. For this single mission, there are at least four different deployed models in operation (see figure 1). New and unique AI models will be developed as the Army fields new capabilities dependent on cutting-edge technologies. Task automation will further increase demand for AI models, supporting software, and their respective hardware platforms.

A rapid increase in the number of useful models will also result from the need to customize existing foundation models. AI model customization, also known as fine-tuning, requires retraining on custom data. Starting from a common base model, models are tailored to achieve higher performance on specific variations of a common task. For example, targeting models will be trained based on local variations in enemy appearance and TTPs, language translation models will be fine-tuned on regional dialects and accents, and weather models relying on physical sensor data will account for known regional patterns. Since there are virtually no limits on the number of variations of a common task, fine-tuning quickly increases the number of models in use.

Enemy-driven model adaptation amplifies the need for a robust model repository in tactical environments. For example, placing tires on aircraft wings or painting geometric designs on ships have been used in conflict to mislead computer vision models.⁶ Models can be trained to recognize these simple tricks, but once deployed, an AI model’s performance, in most cases, depends on the quality and volume of training data. Therefore, if the AI-enabled system senses data from its operational environment that diverges significantly from the training data, the model’s performance will decrease. A change in enemy appearance, behavior, or TTPs can potentially render a model obsolete, or at least degrade its performance beyond reasonable fluctuations. As a result, a robust model repository enables a model to be retrained on current operational data.

One example of adapting a model to enemy appearance is in target detection and tracking. The aforementioned tires on aircraft to confuse detection models did not even require sophisticated tactics. Predictable, single-context sensory input, in this case overhead satellites, enabled enemy forces to deceive the detection models. Clearly, observing the same object from diverse perspectives can overcome this countermeasure and easily detect the aircraft. Multimodal and multicontext AI-enabled systems therefore also provide the ability to counter enemy changes in appearance, behavior, and TTPs. However, when adapting to more sophisticated enemy tactics, retraining on current operational data provides the greatest probability that this new training data (and parameter updates as a result of such training) is as similar to data sensed from the current environment or ingested by query as possible.

In static environments where the query distribution is largely stable across time, a single model need not be updated. However, when an intelligent enemy is actively seeking to evade detection or more generally prevent the friendly AI-enabled system from performing its task, units must have the ability to rapidly retrain their models such that the training data distribution maximally matches data obtained from the operational environment. Similarly, a local repository with guaranteed availability also enables units to revert to previous model versions. Enemy behavior may have previously changed, causing a model update, but then reverted back to its original form. Such enemy patterns require the ability to return to a previously trained model stored in the local repository and deploy it on demand. An AI model repository capable of storing a sequence of model versions along with their associated training datasets is critical to maintain decision dominance.

An Adaptive Unified Repository Framework

In order to store, manage, and train adaptive models, each Army unit must manage its own repository aligned with its mission. The unit’s independent repository needs to be structured and managed using a standardized repository framework (see figure 2). This is the logical consequence of the Army’s “enterprise-i-fying” efforts, which facilitate model sharing and real-time updates.⁷ A unified framework enables units to deploy models received from superior and sister units in the same organization. Further, to enhance operational capabilities, the Army’s decentralized repository architecture will be synchronized in a hierarchy. By maintaining its own repository, a unit has guaranteed access to version-controlled unit- and task-specific models and can also utilize resources available elsewhere in the hierarchy. With the recent momentum in efforts to place AI-skilled officers inside operational data teams together with robust and versatile division and brigade innovation cells, AI-savvy personnel will be increasingly available to oversee and conduct the data curation, labeling, and model retraining and validation process.⁸ These personnel will manage the unit repository to ensure all downtrace units (and any other units throughout the force) are able to access recently retrained and vetted models, which is only possible within a unified repository framework.

Each unit down to the brigade level manages its own repository with the resources they have available (the lowest unit at the AI-edge is likely to shift as technology improves). For example, divisions and corps host what amount to be Army foundation models, generalized for use by units across the division, while lower-echelon repositories maintain models that are fine-tuned to their specific tasks or according to recent enemy activity in the units’ AOs. Each unit maintains a replicated copy of the repository of immediately subordinate units. This facilitates sharing between sister units and provides a backup in case servers are damaged or need to be freshly installed. Not all units will have the same computational needs, tasks, and abilities to update their models, but any updates need to be unified in a system that encourages portability by default. By exercising control over model selection and versioning within its repository, a unit can directly balance its own computational resources against operational needs.

Higher units receiving model updates from subordinate units perform a more robust testing and evaluation process prior to advertising them to other units. Model testing and evaluation must be agile and consist of both manual and automated testing and evaluation processes. Since AI models that update frequently are especially vulnerable to counter-AI attacks and concept drift, testing and evaluation of model performance is necessary to ensure confidence in AI-systems.⁹ All testing and evaluation procedures must balance mission requirements against risk. Models with a larger user base or greater mission impact require a higher testing and evaluation standard before being authorized for deployment to AI-enabled systems. Model testing and evaluation must be transparent to other units, including the datasets and procedures used for evaluation.Using standardized qualitative and quantitative metrics for test and evaluation increases confidence in a unit’s ability to modify and share reliable, trustworthy, and highly performant models.

Each model performs a unique task within an AI-enabled system. The repository stores all versions of the same model, known as the model family, which includes a “model card” explaining the datasets on which it was trained and the performance of each model according to the test and evaluation process. Figure 3 gives an example of how a model card appears in the repository. Each model has a base name, followed by a unit name, which represents its exact position in the task organization in descending order (division, brigade, battalion, company) and concludes with a major and minor version. This naming convention enables the unit to track the source repository from which the model was trained (when receiving another unit’s model update). Metadata for each model version contains information about model and data lineage (all previous model versions from which this version was derived and their respective training datasets) for historical tracking. Additional information regarding datasets, validation procedures, etc., is available via hyperlinks: a description of the expected output, minimum hardware requirements, performance metrics, and evaluation techniques.

As noted, adapting models to enemy TTPs requires a dynamic repository. While some adaptations are automated, many model updates rely on trained personnel who first identify the operational need for a model update and then analyze and validate the updates before they are committed to the repository. These personnel are invaluable to the seamless operation of the distributed and hierarchical model repository framework. Every unit with a repository has trained personnel—for example, machine-learning engineers, data scientists, and network engineers—who can validate, fine-tune, and maintain the necessary combination of software and hardware for the unit to deploy models and improved versions to their mission-critical AI-enabled systems. The trained workforce maintains the repository, validates model integrity, and coordinates with other units.

A unified model repository framework provides a mechanism for modular model exchange across the Army. Similar to swapping compatible ammunition, units can quickly deploy models trained by other units. As units encounter new scenarios, they can update their AI models and share the results with the broader force. Ultimately, a model repository framework empowers commanders to make AI model update and deployment decisions based on trustworthy performance data.

Repository Resiliency to DDIL Conditions

A single model repository stored in a cloud-accessible data center, replicated in different geographic regions, is standard practice for large corporations and other entities requiring high levels of availability. Under permissive and reliably high-bandwidth network conditions, an equivalent cloud-based distributed model repository is the preferred solution for tactical units that possess redundant network transport links such as low Earth orbit and geostationary orbit satellite communications, commercial wireless (5G), and optical fiber if near a point of presence. This is not a realistic approach for all military operations, especially those where the adversary can create austere connectivity environments. Recent reports from the Iranian protests and Ukrainian battlefields document the ability to degrade Starlink low Earth orbit-based internet by disrupting Global Positioning System signals required by ground terminals, demonstrating a potential tangible threat to Army systems.¹⁰ More sophisticated peer adversaries will contest US Army network transport infrastructure and available bandwidth to a much greater degree, threatening reliable connectivity. Therefore, mission-critical models must be stored and available with maximum physical proximity to unit locations as possible to minimize connectivity dependencies across long-haul, wide-area network links. As a result, any unit that may deploy to a location potentially vulnerable to jamming and other forms of network denial or degradation must possess its own locally hosted model repository. As connectivity permits, units must be able to transmit and compare model updates rapidly in a manner that is robust to emergent, dynamic connectivity challenges.

Cloud providers offer almost limitless scalability, and a typical service-level agreement can provide over 99.999 percent of guaranteed uptime.¹¹ However, “no service-level agreement survives first contact with the enemy.”¹² While cloud-based repositories may be fully accessible from tactical environments in most operational scenarios, it is unacceptable for units to be denied access from their mission-critical models on a persistent basis. Units may receive or build new hardware (e.g., drones) in the field, requiring trained soldiers to deploy the current context-appropriate model to the platform. Moreover, in order to adapt to enemy TTPs, the repository framework must be designed in such a way to allow for retraining as close to the deployed models as possible.

Managing models locally requires specific design features in the repository framework. Units need to adapt their models based on enemy actions beyond simply hosting and deploying them. As noted previously, Russian and Ukrainian troops have deployed AI vision model countermeasures in the ongoing conflict. When units identify a model’s substandard performance in various operational use cases, they need a mechanism to trigger updates based on new data. Consequently, locally fine-tuned models expectedly fall out of synch with a higher unit’s base models. When connectivity is restored, higher units store these changes for future evaluation and dissemination while still maintaining the base model for general operational scenarios.

Central to the Army’s model repository is the interconnectivity between subordinate and higher units and network connections between geographically proximal units. Modern network complexity implies that unit preparation for operations in DDIL environments is more complicated than simply disabling all connections in exercises.¹³ Even if the central connection between a unit and a higher command is down, if redundant vertical connections via peers exist, the unit can reroute information through neighboring units. This allows for model updates to be shared with units in the same AO without relying on single links of failure. By planning for DDIL environments, the Army’s model repository framework provides redundancy in model availability and preserves valuable model updates that are often won through tactical experience.

Distribution of Local Knowledge

The primary purpose of a DDIL-resilient model repository framework is to guarantee the availability of successive versions of diverse models to their own units. A secondary purpose is the ability to share new knowledge about changes in enemy appearance, behavior, and TTPs for integration into AI-enabled systems of subordinate, peer, and superior units. Historically, knowledge about new enemy tactics or assets traveled via voice (radio, once available) from one soldier to another or in the form of post-mission reports shared between units. There are two distinct forms of model sharing:

1. A unit transmits the weights (containing new knowledge) of a model to another unit.
2. A unit offloads computational tasks by leveraging resources maintained/owned by an adjacent unit.

As units improve their existing AI models in DDIL conditions, they share updates without having connections to enterprise cloud repositories. Higher units store copies of the repositories from all their subordinate units. Units thus have access to models of peer units via their centralized connection, a form of vertical model sharing. Vertical model sharing depends on tactical connectivity and utilizes distributed systems for delayed asynchronous updates. When direct connections are unavailable, units can send updates to peer units to be routed to higher headquarters when available. Horizontal model sharing refers to sharing models with adjacent units without routing through higher echelons. Horizontal sharing is triggered when a unit advertises a model to neighboring units (a push), and a unit indicates a desire to obtain a model (a pull). The Army’s model repository needs to allow for seamless synchronization between units to share models. Figure 4 shows an example of how a battalion can share models throughout a brigade using both vertical and horizontal model sharing.

An AI model is a combination of a fixed architecture and learnable (trainable) weights. Transferring a model update means transferring the learned weights from one unit to another. Weight transfer can be performed between two units that have either physical or networked connections between them. Simply transferring all weights for each update can be costly as recent models can be on the order of hundreds of gigabytes. Units (or Army policy) must encourage “transfer aware” model design in advance. Practical methods exist that allow for low-cost fine-tuning and model versioning, such as low-rank adaptors or freezing all but a select number of layers.¹⁴ In general, units start with a common base version so that only changes to the weights need to be transmitted. Model repository engineers design plans for updating each model in the repository efficiently and according to operational needs.

Horizontally connected units are able to share network and compute resources, such as hard disk storage, memory, and multiple types of processing, to support model sharing and development. Network resource sharing enables vertical and horizontal transfer of model weights. Compute resource sharing enables optimized AI model operations. The most common scenario in which compute resources are shared occurs when an AO is saturated with targets and requires fast model responses, as illustrated in figure 5. By sending data to neighboring units, per incident response time decreases, targets are tracked across AOs, and model redundancy provides resiliency in dynamic battlefield conditions. Here, the routing process, most likely administered by an AI model, balances latency and workload. Horizontal model sharing provides higher-level units with insight into specific model demands. Dynamic model sharing between units allows for optimal allocation of precious compute resources in tactical environments. Sharing fuel and ammunition is common in high-intensity conflict where contact with the enemy is inconsistent across units. Similarly, units will require a wide variance of computing resources given enemy location, activity, and its resulting perceived threat to friendly forces. The distributed Army model repository framework gives units the ability to share improved models (knowledge about the enemy) and leverage external compute resources on demand (to accommodate spikes in AI model usage).

Challenges and Other Considerations

There are several challenges with implementing an Army-wide model repository framework. Hardware, software, and personnel must be effectively distributed and maintained. The repository framework itself needs to be designed with a unified interface that anticipates future needs without requiring major refactoring. As with all standardized software, increased portability enlarges both the power of sharing and the potential for adversarial actions. These challenges are best addressed with a phased approach, starting with larger organizations and working downward.

The model repository framework must operate at the AI-edge of operations. Currently this edge is at the division level. Graphics processing units—the most common type of processor used by AI models—trained personnel, and funding are largely maintained by division innovation cells. These disparate groups, together with others throughout the Army Science and Technology ecosystem, need a unified framework to store and access models. But the AI-edge of operations is increasingly shrinking. Edge computing, cloudlets, and increasingly accessible model design tools are reducing the technical background necessary for AI operations. Moreover, soldiers across the force are rapidly experimenting and working with AI-powered tools to design new models that solve operational problems.

A centralized repository framework presents unique attack vectors for enemy actors. Manipulated data or models could propagate throughout the Army via horizontal and vertical sharing. Mitigating these risks, like other cyber defense actions, involves both robust validation and testing and the ability to respond and revert models when software-based threats are identified. Lower echelons, manned by personnel with less technical expertise, will be more susceptible to adversarial attacks. Careful monitoring will be necessary before incorporating models or data trained or generated on edge nodes into the larger network. Despite these challenges, a distributed and DDIL-resilient AI model repository framework is essential to provide AI capabilities in contested environments.

The Way Forward

The Army must wrangle and organize the many disparate AI models that power new cutting-edge systems by implementing a distributed model repository framework, empowering units to experiment with, improve upon, and share optimized models. Units are already being bombarded with AI-enabled systems to support operations across the warfighting functions. Without a stable, yet decentralized model repository, units will be unable to manage the diverse set of exquisite models and rapidly deploy them to autonomous systems. Instead, careful thought must be put into the design of an Army-wide framework for AI model management. Designed properly, this framework enables heterogeneous AI models to be adapted to specific use cases by individual units. They must be able to update their models according to changes in enemy appearance, behavior, and TTPs. Moreover, as units submit improved models to higher headquarters, the Army needs a trained AI workforce to evaluate and propagate successful models throughout the force.

A hierarchal distributed Army AI model repository framework allows units to combine knowledge derived from their experiences through decentralized, on-demand model sharing. They can work together to balance bandwidth and latency against operational needs, lending hardware and bandwidth to the fight when and where they are needed most. The ability to update and share models must be integral features of such a system and cannot be retrofitted into the framework ex post facto. To design the AI repository framework, the Army needs to immediately start consulting AI experts, practitioners, and soldiers for testing and integration of existing commercial solutions to generate technical insights and lessons learned.

A key lesson from the war in Ukraine is that AI warfare is only going to increase in both complexity and scale, motivating the Army to make every effort to maintain decision dominance over peer adversaries. This requires the ability to adapt and improve AI-enabled systems as a function of enemy behavior. Units must be able to employ the best model available for the given task, in the specific environment, and with maximum compute resources when feasible. The hierarchical repository framework described herein facilitates these capabilities. A well-designed and unified DDIL-resilient model repository framework enables adaptive AI supremacy, critical for the future AI-enabled Army to fight and win the Nation’s wars.

---

Notes

1. “Anduril’s EagleEye Puts Mission Command and AI Directly into the Warfighter’s Helmet,” Anduril Industries, 13 October 2025, https://www.anduril.com/news/anduril-s-eagleeye-puts-mission-command-and-ai-directly-into-the-warfighter-s-helmet; Colton Jones, “Israeli-Made Computerized Rifle Sight Tested on Heavy Weapons,” Defense Blog, 2 September 2025, https://defence-blog.com/israeli-made-computerized-rifle-sight-tested-on-heavy-weapons/; “Shield AI,” Shield AI, accessed 4 May 2026, https://shield.ai.
2. “The AI Community Building the Future,” Hugging Face, accessed 4 May 2026, https://huggingface.co/.
3. David Vergun, “Battle Looming Between AI and Counter-AI, Says Official,” DOD News, 25 January 2024, https://www.war.gov/News/News-Stories/Article/Article/3656926/battle-looming-between-ai-and-counter-ai-says-official/.
4. SAIC, “DDIL Environment Cloud Management,” US Naval Institute, accessed 4 May 2026, https://www.usni.org/magazines/proceedings/sponsored/ddil-environment-cloud-management.
5. Zita Ballinger Fletcher, “Army Certifies Rapid Anti-Drone Response Team,” Army Times, 11 November 2025, https://www.armytimes.com/unmanned/2025/11/11/army-certifies-rapid-anti-drone-response-team/.
6. Joseph Trevithick, “Russia Covering Aircraft with Tires Is About Confusing Imagine-Matching Missile Seekers U.S. Military Confirms,” TWZ, 13 September 2024, https://www.twz.com/air/russia-covering-its-aircraft-in-tires-is-about-befuddling-image-matching-seekers-u-s-military-confirms; Wes O’Donnell, “Russia Is Painting Dark Stripes on Its Warships to Confuse Ukrainian Drones,” Medium, 11 July 2023, https://wesodonnell.medium.com/russia-is-painting-dark-stripes-on-its-warships-to-confuse-ukrainian-drones-eca37b9e79f4.
7. Leonel Garciga and Deborah S. Karagosian, “A Conversation with the U.S. Army Chief Information Officer,” Cyber Defense Review 10, no. 1 (2025): 17–27, https://cyberdefensereview.army.mil/Portals/6/Documents/2025-vol10-iss1/V10N1_03_Garciga_2025.pdf.
8. US Army Communication and Outreach Office, “Army Established New AI, Machine Learning Career Path for Officers,” US Army, 30 December, 2025, https://www.army.mil/article/289843/army_establishes_new_ai_machine_learning_career_path_for_officers; Sam Skove, “How Innovation Cells in Army Combat Units Are Harnessing Soldiers’ Ideas,” Defense One, 17 August 2023, https://www.defenseone.com/threats/2023/08/how-innovation-cells-army-combat-units-are-harnessing-soldiers-ideas/389518/; Jennifer French, “Interviews: 173rd Airborne Brigade Officers Emphasize Innovation on the Battlefield,” 173rd Airborne Brigade, 13 March 2025, https://www.skysoldiers.army.mil/Home/Home-2/videoid/955206/dvpcc/false/.
9. Gerhard Widmer and Miroslav Kubat, “Learning in the Presence of the Concept Drift and Hidden Contexts,” Machine Learning 23 (1996): 69–101, https://link.springer.com/article/10.1007/BF00116900.
10. Adam Satariano et al., “How Activists in Iran Are Using Starlink to Stay Online,” New York Times, 15 January 2025, https://www.nytimes.com/2026/01/15/technology/iran-online-starlink.html.
11. Camilo Quiroz-Vázquez, “5 SLA Metrics You Should Be Monitoring,” IBM Think, accessed 4 May 2026, https://www.ibm.com/think/topics/sla-metrics.
12. A play on words based on an old military saying, “No plan survives first contact with the enemy,” commonly attributed to Prussian Field Marshal Helmuth von Moltke the Elder.
13. Spencer S. Waters, “Training in the DDIL Environment,” Marine Corps Gazette, September 2022, WE26–WE29, https://www.mca-marines.org/wp-content/uploads/Training-in-the-DDIL-Environment.pdf.
14. Edward J. Hu et al., “LoRA: Low-Rank Adaptation of Large Language Models,” International Conference on Learning Representations (ICLR) 1, no. 2 (2022): 3; Yunhui Guo et al., “SpotTune: Transfer Learning Through Adaptive Fine-Tuning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019): 4805–14, https://www.doi.org/10.1109/CVPR.2019.00494.

 

 

 

Back to Top

July-August 2026