Not all edges looked so nice.
Since I haven't optimized anything I had to work with large error margins.

To make it clear - I use the word "edge" for both cases:
- the while piece edge (contour)
- one side of a piece

therefor I have 1000 contours described with something like 300-600 points per side.

So what about the #complexitytheory ?

Well I had to optimize for space on my working bench. I had about 1m^2 of space and had to move it away if the space was needed for other stuff.

This included:

- dividing whole puzzle into four quadrants (Q1-Q4 in the picture from last post)
- stacking those parts on a hard paper
- creating buckets with puzzle pieces stored in ranges (like 1-50, 51-100 in boxes)
- while working on the quadrant you set focus for the tool which chooses the proper solving direction
- the tool gives you a candidate
- you go to the box
- find the piece (which was very tedious job)
- compare the proposed edges
- mark result (yes/no/next)
- repeat

This all involved a lot of stack and heap work ;-)

So from this point I had 1000 edges stored, I knew where the corners ware. And what now?

Assuming the worst case, where I had to compare all edges. I decided that I wanted to do it only once for each edge pair.

Therefore the main goal of my second #python tool was to compare the edges and hint me with the next probable piece I should try.

Of course I had to go few times back and forth just because I clicked on the wrong button or something.

After many failed attempts to scan or take pictures of the puzzle pieces. I have decided to flip those pieces and work with the grey / blueish background.

I have setup a 30€ webcam placed on IKEA drinking glass pointing down (about 14cm from the lens to the table).

Wrote small #python tool using #opencv ( https://docs.opencv.org/4.x/index.html ) to build an inventory of numbered puzzle pieces. I have stored only the points describing the edges (not the pictures).

So it took me like 3 hours to ingest all 1000 pieces.

I did it once and worked with imperfect data (no optimization, no light adjustment nor camera calibration https://docs.opencv.org/4.x/dc/dbb/tutorial_py_calibration.html ).

I took it as it was, cause nothing is perfect in life ;-)

A thread #introduction :

Recently a candidate in the job interview for my team asked me "what am I doing in my spare time?"

Answer:

If I have enough time I experiment with different stuff (mostly related to computers).

Here one example on solving a #puzzle . Where I had to learn about #computervision , #edgedetection , #complexitytheory (time vs space, wait for it).

The puzzle was given to me as a challenge. "Hey thorkill, you like hard problems to solve..."

It took me over a year to get to solve this puzzle. Thanks to the lockouts caused by COVID-19 outbreak I had time to give it a try. So it all started with this post:

There is still some work to do. But it looks okey-ish. The API endpoint `/timelines/` and using since_id behaves somehow strange. You get "fresh" toots but not the older one. So if you are offline few days, then there is nothing to keep read? Need to investigate.

- since 13.11.2022 I have collected about 2.5k toots/boosts (with some breaks in between)
- there is no way I could keep track of those posts without threads

I have a PoC for Mastodon to NNTP running. Now the timelines make a lot more sense to me.

All you need is:

https://github.com/epilys/nntpserver.py/ and https://github.com/halcy/Mastodon.py

Still work in progress. But it's nice as PoC.