AI Case Study

"As I walked around my office on a recent morning, a female voice on my iPhone narrated the objects I passed. “Brick,” “wall,” “telephone,” she said matter-of-factly. The voice paused when I came upon a bike hung on a wall-mounted rack, then intoned, “bicycle.”

The voice is part of a free image-recognition app called Aipoly that’s trying to make it easier for those with vision impairments to recognize their surroundings. To use it, you point the phone’s rear camera at whatever you want it to identify, and Aipoly will speak what it sees (or, at least, what it thinks it sees) and show the object’s name on the phone’s display. Aipoly runs directly on your phone, so it doesn’t need Internet access to work, and it can identify one object after another as you move the phone around, without requiring you to snap a photo of each thing.

The app was rolled out for the iPhone early this year by a Melbourne-based startup of the same name. Aipoly cofounder Simon Edwardsson says it recognizes images by using deep learning, which is a machine-learning technique inspired by studies of the brain.

The app figures out what something is by breaking an image down into different characteristics, like curves, lines, and patterns (such as stripes), then uses those features to determine the likelihood that the image is a specific object.

Aipoly was able to tell me what plenty of the things are that I can find around my office, though it still needs a lot of work. It can only recognize about 1,000 objects so far; Edwardsson says the company is trying to ramp that up to 5,000, though that would still not be all that many if you consider how many distinct things you come across every day.

And while the app was frequently correct when spotting an object, even when partly occluded (such as headphones slung around my colleague’s neck), there were plenty of misfires, too.

For example, its female voice said “Chevrolet,” then “wheel,” and, finally, “mouse” while the phone was focused on my computer mouse. And the app kept misidentifying a microwave as “AC.”

Users can train Aipoly by typing a correct word or phrase to identify an object that it gets wrong; this data will then be uploaded to Aipoly’s servers and the startup will add descriptions to Aipoly every few weeks or so as it releases new versions of the app, Edwardsson says."