Piqabrick is a project on Kickstarter that aims to deliver a system that will identify parts and minifigures automatically.
As I mentioned on Tuesday, I met the project team, Stefano and Claudio, in Italy at the weekend, who gave me their prototype unit to play with. I've now had a chance to put it through its paces and perform some real world tests.
The results surprised me...
The main physical deliverable of the project is a camera and lighting unit. This one is a pre-production model, so has been 3D printed and is not constructed the same way that the final version will be.
It's exactly 16x16 studs in size and has to be mounted on top of a 12 brick high box in order for the camera to be the correct distance from the base and subject. You can build the box however you want, and in whatever colour you want. I went for a nice simple design which isn't the sturdiest but is perfectly adequate.
The white card at the bottom is provided with the kit and has very fine lines on it which helps the camera auto-focus.
The camera is operated through the Piqabrick website, which at the moment is a work in progress. It currently provides just enough facilities to take photos of parts to add to the database, and to scan parts to test the image recognition. I won't therefore post too many screenshots or pass judgment on its utility yet.
Around 300 parts and minifigs have been scanned into the system so far, which is enough for test purposes but is of course a tiny fraction of the 70,000 in the BrickLink catalogue that need to be done if the system is to be useful.
Project creator Stefano told me that at the moment it works best with compex parts rather than simple bricks and plates, so I thought I'd add a couple of Technic pieces and then see if they would be subsequently recognised.
Adding parts to the system requires taking multiple pictures of them from various angles. The more pictures, the better: I think I took a dozen or so. Of course, you would not need to do this normally because when the project launches the image database should be fully populated.
After a few minutes, I switched from 'add part' to 'scan part and identify it' and it did so without any problems.
These parts are very distinctive so that perhaps wasn't much of a test. I therefore tried with two Alien Conquest soldiers that are identical except for their heads, which, to make it harder, are under helmets.
They weren't already in the Piqabrick database, so I had to find out their BrickLink identifiers and then add them, again, by taking multiple pictures from various angles.
Once done, I asked Piqabrick to identify them. The results were spot-on:
What's more impressive it that it even managed to identify one with the helmet removed, despite me not adding pictures of it in that state to the Piqabrick database.
Of course I have not tested it exhaustively yet but I believe I can conclude that the image recognition works effectively and perhaps better than I expected.
The system is entirely reliant on an accurate library of images of every part and minfigs to be identified and it's going to be some undertaking to provide them for all 70,000 of them by the time the product ships in February. It was not difficult for me to add four for test purposes, but it probably took 1 minute per part which didn't include identifying them in the first place.
Claudio said, in the BrickLink forum, that they aim to have 90% of the parts and figs added by launch and I have no reason to doubt that they can do it, given enough manpower, but it's going to be a huge job!
There has been a lot of discussion about the usefulness of Piqabrick, in comments to previous articles and in various forums, and in particular questioning who's it aimed at. It's probably too expensive for the casual user but having conducted these tests today I think that those who need to identify minifigures and printed parts accurately are likely to find it particularly useful, as are BrickLink sellers who want a quicker way of determining the identity of unfamiliar parts without squinting at tiny writing on them or wading through the BrickLink catalogue.
The project is now about 70% backed with 9 days to go. Pledge your support today to help make it a reality. Who knows, if this is funded, it could be the first step towards an automated parts identifying and sorting system...
If you want me to try some other tests, suggest them in the comments. If I have the parts needed I'll do so.