Summarize this content to 2000 words in 6 paragraphs
If you’ve never woken up to find a raccoon hanging out on your living room carpet, let me tell you from first-hand experience that it’s a surprise best avoided.
One solution, it turns out, is a combination of strobe lights, public radio, and AI.
Back in September, I posted a story and a Ring video from the back porch of our house in Seattle, showing one of these crafty nocturnal creatures prying open our microchip-detecting cat door to get into our house.
The raccoon devoured all the cat food before scaring the daylights out of me when I woke up and got up to see what the commotion was all about. I screamed so loud that the raccoon went running back out the cat door.
Never thought I’d find a raccoon in the living room! @ring camera footage from our backdoor in Seattle early this morning. It found a loophole in the SureFlap microchip cat door. “Intruder mode” now activated. Yikes. pic.twitter.com/oRB9vgiwmz— toddbishop (@toddbishop) September 27, 2024
In the days after that, word clearly spread through the neighborhood raccoon population about the mother lode of grub in our home. One was even fast enough to circumvent the special “intruder mode” that I had activated on the cat door, which locks both sides when a non-microchipped object tries to enter the door. (Fortunately, this one didn’t get in.)
At the end of the piece, I invited readers to help me come up with a tech solution. It was a bit of a reach, perhaps, in an effort to turn my colorful experience into a relevant GeekWire piece. But a Seattle-based artificial intelligence startup took me up on the offer, and solved the problem with relatively minimal effort.
While it’s not the first time someone has used AI to keep an eye on their pet doors, the effectiveness of the approach shows how much progress some startups are making in developing and applying AI to real-world problems.
The company that got in touch after the publication of the original piece is Groundlight, co-founded in 2020 by CTO Leo Dirac, former Amazon senior principal engineer; and CEO Avi Geiger, former Microsoft principal architect who was co-founder and CTO of homebrewing startup Picobrew.
Groundlight co-founders Leo Dirac, left, and Avi Geiger. (Groundlight Photo)
Groundlight, which raised $10 million in seed funding last year, is backed by investors including Madrona Venture Group, Greycroft Partners, Founders’ Co-op, Flying Fish, Ascend, and Essence VC.
The company’s technology combines computer vision with human review for AI model training and image recognition in industrial and commercial settings, using a small device that it calls the Groundlight Hub.
The idea is for businesses to easily build and deploy computer vision systems tailored to their specific needs, without requiring expensive, high-powered hardware or extensive machine learning expertise.
The possibilities are basically endless. For example, customers might use Groundlight’s technology to identify anomalies in industrial machinery, detect leaks, monitor the status of inventory on a shelf, make sure warehouse aisles are clear, count the number of empty spaces in a parking lot, or ensure that dumpsters aren’t obstructed on trash day.
Detecting raccoons? Not a typical use case. But definitely possible.
At least, that was the idea floated by Julia Gall, a Groundlight product marketing manager, in an email after my story ran. She explained that she’s from a family of cat lovers, and her mom shared the article with her. The concept was to train the computer vision model to distinguish between our cats and raccoons.
After an initial call, and some input from Groundlight’s founders, she and Groundlight software engineer Brandon Wada visited my house in late October to set up the system. He installed a Groundlight Hub, a small orange device on our kitchen counter, connecting wirelessly to a camera on the back porch, pointed at the cat door.
Brandon Wada, Groundlight software engineer, tested the system with cardboard cutouts after the initial installation. (GeekWire Photo / Todd Bishop)
The AI was given a simple question: “Is there a raccoon?”
Groundlight’s approach is based on a confidence score, not a simple yes or no. If the confidence in the answer of “yes” exceeded a preset threshold (90% in this case), the system was set up to activate a smart plug.
The initial hypothesis was that flashing lights might be enough to scare the raccoons away. So the solution that Groundlight came up with was a light box, plugged into the smart plug and facing the cat door, with a strobe effect so bright that I could see the patterns in my eyes for more than a minute after it went off during a test.
Of course, the key was to make sure that this strobe light was unleashed on the raccoons, not on our cats. I was a little concerned at the outset, given that our cats, Snuggles and Emily, have white, gray and black coats, with tails that get bushy when they’re cold or scared. This seemed to me like a recipe for mistaken identity.
While the downsides of a false positive wouldn’t be catastrophic, they also wouldn’t be pleasant — especially for Emily, the more skittish of the two — so trusting in Groundlight’s approach required a leap of faith.
Groundlight’s system distinguishes between a cat tail and a raccoon tail. (Groundlight Image, click to enlarge).
Initially, during the training period, the system incorporated human review to confirm the presence of a raccoon. This rapid human-in-the-loop process, available 24×7 via Groundlight’s global network of reviewers, is a big part of the company’s secret sauce.
In the days after the installation, when the raccoons would show up, the AI initially wasn’t confident enough to exceed the high threshold that had been set to activate the switch.
However, as the days progressed and the system gathered more data, it eventually learned to identify raccoons on its own with enough confidence to trigger the smart plug and activate the bright, flashing light.
(See the Groundlight dashboard from the raccoon implementation here.)
Shortly before 6 a.m. on Nov. 8, a raccoon strolled up to the back door. Within a few seconds, the bright light was flashing rapidly, lighting up our back porch with a crazy strobe effect. And the raccoon was totally unfazed, staring directly at the light before strolling away casually. If he could have shrugged his shoulders, he would have.
OK, so bright lights aren’t enough. After some more brainstorming, the Groundlight team came up with a new idea, sending me a cheap radio to add to the setup, along with the flashing light.
I tuned it to KUOW public radio, setting it to blast at full volume whenever the smart plug turned on, under the theory that human voices would be more likely than music to scare the racoons.
At 12:56 a.m. on Nov. 20, the combination of the AI, the strobe light and BBC World Service did the trick — recognizing a raccoon and activating the radio and light. Our cameras captured the raccoon scurrying off the porch, clearly not knowing what in the heck had just happened.
“Success!!” read my email to the Groundlight team, which had about 50 messages at that point.
Proving that it wasn’t a fluke, the system has worked on multiple occasions since then, scaring raccoons away. And even better, the AI has never incorrectly identified our cats as raccoons or flashed the light on them.
I jokingly asked the Groundlight team where solving my raccoon problem ranks among the engineering accomplishments in their careers. Geiger, the company’s CEO and co-founder, said it actually ranks relatively high, given that it’s a such a clear proof point for the company’s strategy of democratizing computer vision.
My relatively trivial problem (in the scheme of things) normally wouldn’t have warranted the investment required to create a custom model. But the company’s platform is flexible enough to be easily applied to a wide variety of situations, without extensive machine learning expertise, or expensive hardware.
“That’s a huge accomplishment,” Geiger said.
The architecture of the raccoon detection and deterrent system. (Groundlight graphic, click to enlarge.)
This was illustrated by the fact that the biggest challenge wasn’t identifying the racoons — the AI had that problem completely solved within a week or so — but in coming up with the best way to scare the raccoons away.
Two papers from Groundlight researchers and engineers have been accepted for workshops at the Neural Information Processing Systems (NeurIPS) conference, including one for “Fine-tuning Vision Classifiers on a Budget,” and another, not yet published, on integrating human judgment into machine learning models.
Currently, implementing Groundlight’s solution requires using the Python SDK and writing a minimal amount of code to build an application. But one of Groundlight’s projects in the pipeline is a self-service tool that will allow non-technical users to configure and deploy its computer vision solutions without writing any code.
“It’s going to be a completely no-code solution where you just configure it to find your camera, ask the questions, set up what kind of alerts you want, what kind of conditions, and anybody can do that with very little technical expertise,” explained Dirac, the company’s co-founder and CTO.
As for me, with apologies to Seattle’s raccoon population, I’m already sleeping a little easier knowing that AI is keeping an eye on my porch.
