‍

‍

Computer vision models are increasingly being adopted across several industries and verticals to optimize workflows and solve challenging problems that in the past have been cumbersome and time consuming. Whether it is self-driving cars, or factory robots with advanced camera equipment, giving machines the ability to see and comprehend their surroundings is a powerful tool for boosting productivity and bringing new products to market.

‍

One computer vision use case that we’ve likely all encountered at one point or another is face recognition. Modern cell phone cameras use facial recognition to detect where the faces are in the shot in order to ensure they are in focus before capturing the picture. You may have also encountered social media platforms that automatically recommend tagging your friend in a photo you’ve uploaded. In this case the model is not only detecting that there is a human in the photo, but is also able to correctly identify who it is based on previous tags. The security industry has also utilized facial recognition systems as a tool monitoring checkpoints at airports and other establishments such as businesses with restricted areas.

‍

Celebrity Recognition Use Case

Another area where facial recognition algorithms have been deployed successfully is celebrity recognition. Celebrity recognition models use computer vision to detect and identify celebrities in both images and video data. Companies—mostly in the media space—use these models to easily search for footage or pictures of specific people they are looking for. With the advent of inexpensive and high speed internet, and greater cloud storage capabilities, video data has seen an exponential growth. The problem with video data is that it is not as easily searchable and filterable as text and to some extent even images. Celebrity detection models allow companies to filter out scenes that contain certain famous people in them.

One use case of this technology could be for online streaming services to automatically tag all the movies that an actor or actress is in. While this may seem like a problem that is more easily solved by simply scraping existing publicly available databases of movies and cast lists, consider video data that is more unstructured. News programs and interview shows host celebrities to speak about their latest projects and to get updates on their careers. Celebrity recognition models can automatically tag the celebrities in the interviews so that the data can be filtered by actor/actress.

‍

Celebrity recognition models can also be useful for sports broadcasters. Sports media companies often need to show highlights of a certain player during a specific segment they produce. In the past, the task of labeling the footage data was assigned to teams of people that would have to go through the footage and tag each clip with the player’s information. This is a timely, inefficient and expensive procedure. Today, media companies are able to use celebrity recognition to solve this problem faster and cheaper. Celebrity recognition can also be used for detecting what celebrities (non-players) are in the stadium or arena at the time to showcase them on television as is often done during live games.

‍

Challenges with Facial & Celebrity Recognition

One challenge with deploying systems such as celebrity recognition models has been that it requires teams of machine learning engineers which media companies may not have access to. That problem has been solved over the last few years by large tech companies such as Google and Amazon which have provided public APIs through their cloud services for celebrity recognition. This has made access to these models much cheaper and far more easily available.

Other challenges with both facial recognition and celebrity detection include factors such as the pose of the person whom you’re trying to identify, and blockage of their face in the video or photo. The resolution of the photo can also play a role in the model not being able to accurately identify the person in the photo. The tasks of these models can also become more complex when there are several people in the frames of the photo that it is trying to identify people in. This becomes even more difficult in the sports analytics use case where there could be hundreds of people in the background, sitting in the arena.

As with any machine learning problem, having access to good, representative data is key to having a model that performs accurately in the real world. While everything should be done to have a strong dataset in the pre-production environment that can account for all of these factors, it is also important to realize that the performance of the model will decay over time. The solution is to monitor the performance of the models in the production environment, as the model is performing the task it has been given. Identifying outliers and introducing them to the model’s training dataset will improve the accuracy of the model in an efficient and cost-effective manner, avoiding bloated datasets and reducing the feedback loop that it takes to improve computer vision systems.