DeweyVision is transforming post-production in the Media and Entertainment industry, working with top Hollywood studios to streamline their workflow. DeweyVision’s cutting-edge Post-Production AI Copilot is designed to simplify how film footage is assembled, conformed, and archived—turning days of tedious manual work into mere minutes. By eliminating mundane tasks, studio professionals can dedicate more time to what truly matters: bringing creative visions to life.

Post-production projects often involve finding a needle in the haystack – i.e., given a particular image, find similar frames or scenes by searching through petabytes of raw movie footage involving different cameras or with different visual effects and transformations. Finding frames accurately and being able to seamlessly integrate them in the final release requires tracking them through every stage of post-production.

The benefits of using AI vectors to search video content

Manually searching through visual data can take hours or days, which is why DeweyVision makes video content trackable, frame search accurate, and fast—taking seconds instead of hours, thanks to AI vectors (called Dewey Vectors in DeweyInsights). AI vectors are a mathematical representation of unstructured content such as the individual frames that make up a movie. DeweyVision processes each frame in the raw movie footage, creates vector embeddings for each one, and stores them in a vector database. This allows DeweyVision to quickly search the vector database using simple math that’s built into the database to find similar vectors, and hence similar video frames. To increase scale and further improve performance, DeweyVision is now moving to use Oracle AI Vector Search in Oracle Database 23ai, which enables fast AI-powered semantic searches across various data types by introducing native vector support, optimized indexing, and leveraging NVIDIA GPUs to accelerate vector embedding generation..

“Oracle Database 23ai with AI Vector Search can significantly increase Dewey’s search performance while increasing the scalability of the DeweyVision platform”, said Majid Bemanian, DeweyVision CEO. “Using NVIDIA GPUs to create the vector embeddings that we load into Oracle Database accelerates the speed at which we can ingest new data, while Autonomous Database and the converged capabilities of Oracle Database 23ai will help reduce our operational costs as we grow and open new opportunities. We believe that the combination of DeweyVision, Oracle Database 23ai, and NVIDIA GPUs running in OCI will help us achieve our goal of becoming Hollywood’s data warehouse.”

 

DeweyVision search
Figure 1 Dewey Vision Discovery and Search

Challenges with stand-alone vector databases

Using vectors has been a game changer for DeweyVision, dramatically reducing the time and effort for studio post-production. They initially stored relational metadata in PostgreSQL and used a combination of Qdrant and FAISS vector stores for similarity search. However, their initial multi-data store approach came with its own set of challenges:  

  1. Managing and synchronizing disparate data stores: Managing multiple data stores involves dealing with governance, protection, data inconsistencies, maintenance, and other aspects separately for each store. Layers of code and tools need to be implemented and maintained to achieve seamless data management, which adds to the overall complexity.
  2. Inefficient complex queries: Hybrid searches involving both attribute filtering on metadata and similarity search on the vectors are complex and inefficient since the data is spread across different data stores that need to be accessed separately.
  3. Poor performance: Traditional vector stores keep vectors in memory, but performance degrades significantly when there is insufficient memory. For example, a two-hour blockbuster movie. It comes from 500 hours of raw footage, which is converted into 43 million vectors at a 24 frame-per second rate. In some cases, the number of vectors can reach 100 million, requiring terabytes of memory—far more than what is typically available.
  4. High cost: Maintaining multiple data management systems to support both relational and semantic search adds significant costs and complexity, and can further compromise security and data consistency due to translating and communicating across the disparate systems. All of this overhead drives up costs. Additionally, DeweyVision was using a fixed GPU cluster to process their vector embeddings, which increased their costs when high throughput wasn’t needed.
  5. Protection: Protection is an essential aspect of movie production today.  Film production requires separate teams working completely independently of one another, often not aware of the other aspects of the movie in production. This not only introduces an additional layer of complexity for the post-production teams as there is no single stream of footage to work with, but it also means the content produced by each team must be adequately secured and inaccessible by other teams.                        

Oracle Database 23ai and AI Vector Search benefits

Looking for alternatives to solve the above challenges, DeweyVision approached Oracle to test their solution on Oracle Database 23ai with  Oracle AI Vector Search. After several months of testing, DeweyVision found that Oracle Autonomous Database on Dedicated Infrastructure solved their problems. Let’s take a closer look at how it addresses their challenges.

  1. Converged Data Architecture: Oracle Autonomous Database’s (ADB) converged data architecture unifies structured and unstructured data, so there is no need for separate databases for the metadata and the vectors. In Oracle Database 23ai, Oracle introduced a new VECTOR datatype, allowing DeweyVision to store the vectors alongside the metadata and keep them transactionally consistent. ADB is also a fully managed service, meaning Oracle automatically maintains the system for DeweyVision.
  2. Simplified queries: With the video frame metadata and the vectors in the same database, it is no longer necessary to write application code that separately accesses the different data stores and combines the results. Oracle AI Vector Search enables similarity and metadata searches to be performed together in a single SQL statement, reducing the complexity of their application.  
  3. Higher performance: DeweyVision benchmarked AI Vector Search on ADB against their previous solution. ADB proved to be up to 7X faster than DeweyVision’s previous solution when searching on millions of vectors. Oracle’s vector indexes and Exadata’s AI Smart Scans and high availability deliver mission-critical AI at any scale by transparently offloading key operations to Exadata smart storage and automatically parallelizing and accelerating searches.
  4. Lower cost: ADB’s built-in automation enabled DeweyVision to reduce time-consuming and costly manual management tasks, while its built-in AI Vector Search capabilities helped eliminate the need to translate, communicate, and synchronize data across disparate systems. In addition, by using flexible OCI GPU clusters accelerated by NVIDIA, DeweyVision were able to accelerate frame vectorization and lower costs by scaling GPU cluster size to meet changing requirements.
  5. Built-in protection: With ADB, security and separation of duties are built in, allowing DeweyVision customers to secure and protect each team’s content by making it inaccessible to other teams without requiring additional software components.

DeweyVision’s conclusion

DeweyVision concluded that using Oracle Autonomous Database on Dedicated Infrastructure with Oracle Database 23ai is their preferred choice to make their vector search service faster, more scalable, more accurate, and less expensive than their previous solution—with the additional advantage of getting more functionality and governance for the vectors. Furthermore, DeweyVision was able to move from having multiple types of search and storage in favor of a unified database. The result is a simpler deployment, a faster product, and less room for duplicates, inconsistencies, or gaps.

DeweyVision has extended its deployment options to include Oracle Autonomous Database Dedicated with Oracle Database 23ai.

The combination of Oracle Autonomous Database Dedicated with Oracle Database 23ai  for frame search and NVIDIA GPUs for vectorization is taking DeweyVision to the next level!

 Learn more