Improving our Amazon Picking Challenge baseline with data

Now that we have a baseline (you can see initial results here), we want to see where it struggles to get an idea of how we're going to improve it. Nothing sensitive here, so if you're looking to scoop our amazing work, move along.

To make these improvements, we need data!

Originally we were going to capture data by manually moving objects in front of our RGB-D camera, but had some spare parts from the Mobot project and decided to build a hacky turntable.

First, some basic materials. We had a 12"x12" acrylic tile and a disassembled lazy susan base.

Note to self: you don't need a tap wrench when you have a power drill and can turn the chuck by hand.

And it's all going in a box.

Which is a little too big, relative to the actual Amazon product shelves.

But we're going with it.

Not too bad!

Using this system, we were able to capture several dozen images for a few test objects chosen from last year's set. Here are some examples:

Using the same simple algorithm from before to distinguish between the Expo dry-erase marker and The Adventures of Huckleberry Finn:

Ok, so these were two pretty easy objects (rectangular, not very reflective or transparent surfaces, lots of words and images to pick up visual cues). We capture some harder objects and will be testing those soon.