The A9 Problem

So I am off to Los Angeles for the Microsoft Professional Developer's Conference (PDC). It is worth noting that I am a tea junkie, occasionally called a "tea punk" by my friends. Whenever I travel I like to find a good place to have a cup of tea in the morning before the conference starts up. The last time I was in LA I went to a fine little establishment called "The Coffee Bean and Tea Leaf." They had wonderful tea, delightful scones and friendly staff, the perfect morning combo. I wanted to find out if the store was still there and if not where else I ought to go for tea. This is where A9 comes into the picture.

A9 is an Amazon company whose mission statement sounds more like bragging than corporate direction. People at work are unbelievably excited about A9 and the possibilities of storefront mapping. If you aren't familiar the idea is to have a truck with a GPS and a camera or video recorder fastened to the top of the vehicle. As the truck drives down the street it takes a constant stream of pictures and associates the images with a location from the GPS. There is fancy software in the van which makes up for the "urban canyon effect", which prevents a GPS from accurately reporting position. The pictures taken at, or near, a location are then tied with the location of a business by geocoding and viola for a given business you can show a picture of the storefront.

As a cynical guy, I found the whole idea totally unworkable and so I tested A9 a few weeks ago. In two words "it sucked." I searched for Chinese restaurants in New York City. Of the top ten results, nine of them had pictures listed. I examined each of the nine pictures and found that six of them, the top six, showed vacant lots, sides of trucks or the wrong location. For instance, the first result is Kam Sung Chinese Restaurant with a picture depicting the spray-painted side of a building. The actual restaurant is four photos to the right and even that picture is partially blocked by an SUV.

This time I searched for "tea" and "figueroa st, 90017" (I had to use the ZipCode since "figueroa st, los angeles, ca" failed). The results page included a reference to 'The Coffee Bean and Tea Leaf,' in the lower right-hand corner. The picture it showed was the intersection four pictures to the right of the store. If I did not know what I was looking for the picture would not have helped. I noticed, however, there was a series of checkboxes below the image with icons allowing me to "walk left" or "walk right."

Since I had been there before I knew it was left a few steps, so I clicked "walk left" and found another A9 problem. The direct shot of the store was bathed in bright sunlight for the right hand 50% of the picture. Though I knew the store was there, the sign was hard to read because the camera had been overwhelmed by the sunlight reflecting off of the building. This is one of the problems encountered by aerial photography. The cloudcover, sunlight level and angle of the sun are all vitally important. You can see this mistake in the aerial photography of Los Angeles in Google Maps (and Google Earth). The poor sun angle of the imagery in Google make the buildings look like the are leaning over, an optical illusion created by the extreme shadows. In the case to the right, however, the problem is created by the sun angle being two low relative to the truck and lighting the building. If the sun where more overhead the shadows would be on the ground rather than on the wall of the building. I suspect the A9 folks are probably being more careful about driving only when the sun is directly overhead, particularly in large cities.

A9 did manage to redeem itself with what appears to be a new feature. I noticed a row of checkboxes below the image list, when I hovered the picture a piece of text showed beside the box: "Best Image?". I selected the picture one more to the left, which depicts the store in 80% shadow. The sign was more readable even though the store was not filling the photo region and the doorway was hidden by contrast. I selected the checkmark and A9 registered this particular picture as best representing the store. This allows people using the site to fix the geocoding by marking the appropriate picture, a good way to handle data quality issues. It is susceptible to misuse since a competing business could shift away from the proper image by checking a box which does not properly depict a location. I imagine A9 will work that out down the road.

In the end, my view on A9 is interesting but not useful. With more experience, better geocoding and a better corrections system A9 might become useful in the future, but that is probably a bit further in the future.