You may also be interested in:

Media Asset Management systems are a fundamental element of the broadcast media supply chain. Although systems for news, delivery, production and archiving have many shared core technologies, each has evolved its own nuanced functionality to fit workflow demands. Many vendors offer different applications accordingly. This guide explores the fundamental principles of Media Asset Management system design and implementation with a firm eye on production and archiving... but many of the standards and best practices discussed here apply across all MAM systems.

Media Asset Management systems can become more complex as they evolve. The key to achieving long term success lies in a solid understanding of the underlying technology, design principles and terminology. Careful consideration of the scope of the system and following established standards and protocols which will streamline future expansion is critical.

Managing assets becomes a bigger and more complex task as the scope and scale of ingested content increases. AI provides leverage when wrangling these huge quantities of media assets, but knowing what happens under the hood of an asset management system is key to deploying it successfully.

Integrating Third Party Products With Custom Extensions

Hardware engineers will be familiar with introducing AV glue into their architecture designs. This integrates devices from different manufacturers so they can exchange content and metadata. Whilst both devices may state that they adhere to a common standard, conflicting profiles and levels or sub-optimal implementations can prevent it working. The AV glue rectifies that.

It is the same with software-based solutions in IP workflows. Buying off-the-shelf content management tools will rarely provide everything you need in the longer term. Additional asset managers, external database tables, ancillary software tools will often need to be integrated with the proprietary systems. Software glue does the same thing.

From past experience with integrating multiple closed systems, the lowest risk approach treats a third-party system as read only. Never try to alter the internal structure of a proprietary database. Linking to it via queries and adding relational joins to supplementary tables to ONLY read data does no damage. Sometimes proprietary systems will provide an API to support user data input. Using that would ensure the internal data structures are undamaged.

Managing Multiple Asset Stores

Choosing the optimum asset management software is going to be different for every organization. To start with, a single storage container may not be the best solution. Choosing an asset manager that is very good for wrangling video clips may not be optimal for images for example. Collating sound effects is different to managing music tracks. Managing collections of classical orchestral music requires several additional metadata fields that are not needed for popular music. Classical material is also usually organized by composer rather than the recording artist.

Perhaps you could structure the storage systems around different types of media to start with. Then add a supervizing layer on top that allows searches to be federated across all of them. If the search results can then be collected into a basket to be forwarded to the NLE as a project, the system approaches the capabilities that Final Cut Server provided. Some of these media types could be split into multiple separate containers for each sub-category.

The librarian layer sits above all of the different repositories and provides a means to search for related assets and browse the collection as a whole. Where assets are designed to be used together, that grouping is managed in the supervizing librarian layer.

This list is not exhaustive and other content types will extend this basic concept.

Import-export Issues

Moving metadata from one repository to another should always preserve all of the data so the transaction can be reversed to send the media back to the source system without losses.

This requires that the target metadata repository has a corresponding field for every property in the source metadata repository. The target repository could support extra fields as well, but those additional fields cannot be transported back to the source.

Any loss of data during a transport process is detrimental and could be very subtle. Here are some examples:

The Impact Of AI On Asset Management

Media Asset Management is an area where AI has a lot to offer. The industry is already being transformed by the use of Machine Learning. AI can certainly offer help when ingesting content and provides more assistance when searching.

We will be getting into the subject of AI more deeply in another series. Perceptive AI covers a whole spectrum of activity from simple deductions to advanced pattern recognition. For now, let’s examine some simple requirements for Perceptive AI tools that extract metadata or enhance the results of search engines.

AI does provide some useful benefits when extracting information logged by a camera for storage in the metadata repository. However, is this really a sophisticated AI process or just simple transformations of embedded metadata? It is beneficial either way and delivers information that location crews rarely input properly or completely.

We might expect professional digital cameras to at least log the information that a mobile phone records and a lot more besides:

• Date of shot.
• Time of shot.
• GPS location of shot.
• Directional orientation of camera (if the camera is smart enough).
• Azimuth (tilt of camera if it is smart enough).
• Altitude.
• Camera mode.
• Exposure settings.
• Shutter details.
• Lens characteristics.
• Focus.
• Resolution.
• Frame-rate.
• Color model.
• Sensor type.
• Imaging sensor response curve.

Some of these values might need to be logged on a frame-by-frame basis when shooting in handheld mode or with Drones and Aerial photography.

Other information that AI techniques can help extract concerns the cast and crew participating in the shoot. If AI has access to all of the documentation related to the pre-shoot planning, much of that might be inferred.

Each individual field in the metadata database will need a separate AI prompt to extract the specific information required. Similar perhaps to a database query in a conventional system. This would also be working with a Small Language Model (SLM) based on a contextually focused set of training data.

AI is certainly useful when logging points of interest in sports footage. For example we have known for a long time that the compressed video stream bit-rate bursts and the audio amplitude increases significantly when a goal happens in a football match. AI should be able to detect those events in video without any difficulty. They could made available to the presenting commentators and pundits for near instantaneous action replays.

Taking Perceptive AI Results At Face Value Can Be Risky

Machine learning can deduce things about the content but there are some interesting issues with the analysis that need to be taken into account.

At a demonstration of the IBM AI work in progress, I saw an image recognizer deduce the correct name of a celebrity with a confidence level of about 90%.  It also simultaneously deduced the gender of that celebrity with only a 65% confidence. Given that the celebrity gender was already well-known this was odd.

The two ML pattern recognition results were arrived at completely independently. Introduce an additional layer of logic to rank the results according to the confidence level. Then if possible, override the lower confidence results by inferring them from higher ranking ones.

I don’t think this is yet an exact science. Perhaps that ranking can be automated, but some capacity limited human moderation might be necessary.

This is all analogous to the quality of speech to text recognition systems that are not always completely accurate. The fundamental question is, “What level of inaccuracy is tolerable” because that translates directly into the quality of our metadata when AI is generating it for us. For the time being, it could be checked by human operators to measure the accuracy.

Relevant Standards & Other Resources

ISO has codified asset management principles and published them as the ISO 55000 series. This breaks down as:

• Physical asset management.
• Infrastructure asset management.
• Fixed assets management.
• IT asset management.
• Digital asset management.

There are not very many helpful books covering Asset Management design. This one is well respected and written by an experienced media librarian who has worked for major UK broadcasters.

Conclusion

There has always been some tension between the counter arguments for building your own system vs. buying an off-the-shelf solution.

A bespoke solution may be more secure since it is a one-off design, and any potential security intrusion exploits will not be in the public domain. The internal workings and attack surface of a commercial product may have well known intrusion exploits. All systems demand ongoing security mintenance; with a bespoke system that falls to you, where an off-the-shelf system should benefit from vendor maintenance.

Bespoke systems can be very closely tailored to your specific needs as opposed to compromising your workflow by forcing it to conform to the functionality of a proprietary design.

There is always going to be some ‘glue’ work to be done when bringing up an asset management system using third-party systems. To start with they can be operated independently. At some point adding a layer of bespoke code to control and aggregate the search and extraction process is beneficial.

Careful configuration is also important and it may be possible to write plug-ins that assist with integration. Output bins might be configured to point at watch folders used by other applications. Conversion tools can be interposed to resolve file formatting conflicts using the same technique.

If you use AI to any great extent, keep the human in the loop until you have verified that the AI is working correctly. Perceptive AI may misinterpret the content occasionally but this is probably not as serious as the hallucinations that Generative AI suffers from. Keeping track of how well it is doing with human quality checking, and scoring the results will lead you to the necessary improvements with a little statistical analysis. If you find some problematic imagery, feed it back to the AI engineers to analyze so they can improve the recognition performance to deal with edge cases.