It's been quite a while since I've had to wear a tie for work. I haven't worn a tie regularly for work since my days at Vanderbilt University when the directive came down that managers would wear ties. I was well prepared with a good selection of bright collection ties. So why am I bring this up now?
Over the past few weeks, I have been automating my work-from-home setup. For example, my east-facing home office gets a bit dark after noon so I flip on some lights if I happen to be standing at my desk. I'm usually there so that one less manual step. Yes, I could probably buy something for that task, but building it is so much more fun. Since one of my pipe lines contains an object detect step, I decided to log all of the items detected. I planned to go back through that list to see what was detected and brainstorm any automation I could wrap around it.
Well, a "tie" was in that list of recurring detections. I haven't worn a tie in years, yet here we are. I started looking at the images associated with the log entries. There are definitely no ties in any of the images. What is causing this false detection?
After going though the images and comparing them to the previous and next days images, I suspected that I knew the cause for the false detection. I made a quick adjustment to my pipeline to add some bounding boxes to my object detections. I was correct. The inverted triangle of my black t-shirt peaking out from a light colored sweater or hoodie was being interpreted as a Window knot. The shape is a near match for a perfect Windsor knot, but the blade of the tie is missing. What does it all mean?
In short, understanding your algorithm is critical. Blindly trusting a bit of tech without understanding its limitations can lead to disastrous results. While the side effects of my WFH automation erroneously detecting tie are minimal, a false detection in an enterprise environment could range from simply embarrassing to downright financially impacting.
Now wondering if it's worth the effort to add gestures to my WFH automations...