Why This Matters

Computer-vision solutions are often developed as isolated projects: one dataset, one model, one problem.

That approach can work, but it makes expansion slow and expensive. Every new application may require new training, new infrastructure and a new integration path.

Our research explores how shared visual foundations can be reused and adapted across multiple applications.

Research Direction

Instead of treating every vision task as independent, we are interested in architectures with a common visual foundation and specialised adaptation layers.

Conceptually:

Shared visual foundation → task-specific adaptation → application logic

Possible adaptation mechanisms include:

• Fine-tuning.
• Lightweight adapters.
• Task-specific heads.
• Prompt-based or multimodal conditioning.
• Shared embeddings.
• Reusable reasoning components.

What We Are Exploring

Shared representations

Can one model learn representations that remain useful across several operational domains?

Efficient specialisation

How much application-specific training is required to adapt a common model to a new task?

Model families

When is it better to maintain one general model versus a family of related specialised models?

Reuse beyond perception

Reuse can also happen above the model layer, through shared reasoning, event schemas and deployment patterns.

Why Reuse Matters

A reusable model strategy can potentially improve:

• Time to build new applications.
• Consistency across solutions.
• Dataset efficiency.
• Model maintenance.
• Deployment standardisation.
• Hardware utilisation.

It also aligns with the broader OCCIVUE direction of enabling new visual solutions to be assembled from modular capabilities.

Research Challenges

Reusable models must balance breadth with precision.

A model that is too general may underperform a dedicated specialist model. A model that is too specialised loses the advantages of reuse.

Our research therefore focuses on finding practical trade-offs between generality, accuracy, adaptation cost and inference efficiency.