As security cameras have become ubiquitous they have become problematic to manage. Who is going to monitor the dozens of video streams from a single facility? Who is going to respond to the plethora of false alarms that come from simple motion detection? How will you overcome the fact that humans are inherently terrible at monitoring for rare events? In the future an AI will monitor video streams and only alert a human when it has detected specific troubling behavior.
Umbo Computer Vision is a San Francisco startup with offices in Taipei and London. This global reach allows them to tap into the mix of hardware, cloud and AI skills they need to build a new generation of smart security cameras. These are AI enabled devices that directly address this need for continuous monitoring without false alarms.
Here are the key takeaways from my recent talk with Umbo’s CEO Shawn Guan. Our conversation focused on the AI side of their product development.
Making the leap from lab to reality
Shawn emphasized that optimizing algorithms in the lab only gets you so far, the real challenges are in the deployment and practical application of those algos.
Need for Edge AI is here today
Umbo is one of the few companies that have already put those ideas into practice with very intriguing results.
Small models can be beautiful
Sometimes the size of the models we create matters. Google had an interesting example of this size sensitivity in the “learned index” research they have done.
Umbo with their need for custom models executed at the edge is an example of size sensitivity in a real world setting. This is a good reminder that pure model performance is not always the best measure of utility.
Mass customization of models
We are starting to realize that it doesn’t always make sense to have one model to rule them all. This is especially true in IOT use cases where you want to make each smart device fit its environmental conditions and you are sensitive to model size.
Umbo is distinguished by the emphasis they put on automating the mass customization of models, so that each device is trained with just the right mix of shared historical data and device environment specific data.
Agile model update cycle
Some things you know intellectually from the beginning but don’t understand viscerally till after you take your AI to production. For Umbo one of those lessons was the importances of agility and having a short cycle from identifying an issue to getting an updated model pushed to production.
If a customer has 100 alarms each windy night because there is a branch that uncannily resembles a burglar, they don’t want to hear about how your devops procedures will delay getting them a resolution, they want the resolution ASAP.
Visual explanations for visual use case
Umbo’s system uses visual images (such as the one above) to “explain” why alerts are triggered. Based on their first hand experience they have implemented some nice touches to these explanations. For example notice in the photo that the person scaling the boundary wall is highlighted with a tight fitting shaped outline instead of just approximate rectangle.
Don’t always need full explanation
There are costs, in terms of development team time and system performance, associated with explanations and those costs are higher the more rigorous and complete you make your explanations. The team at Berkman Klein Center proposed one framework for deciding if an explanation warrants the cost. However, each application needs to answer that question based on their own circumstances. The Umbo visual explanation does not the meet the standards of algorithm transparency and global model interpretability. However, it does meet their audience specific needs and that is what really matters.
Easy to assume data more general than it is
Like many AI based startups, when Umbo first launched they had a need for training data but no operating history to give them the opportunity to gather that data. Therefore they did the natural thing of priming the pump with publicly available data sets, and like many others they discovered that seemingly comprehensive training data sets can often fail to provide the variety needed for good generalizability. Umbo with its growing installed bases is now past these issues, however other early stage projects need to keep this lesson in mind.