I searched squares starting from each corner. This got me AC.
Here is a test for which my program fails:

5
1 1 4
1 1 2
1 4 2
4 1 2
4 4 2

You will notice that if you try to reconstruct the squares
starting from the corner images (the way my implementation
does), you will find the small squares of width 2 but not
the larger one (whose corners have been completely hidden).
However, the edges of the larger square will have remained
uncovered by that solution.

You might want to consider adding a similar test
to the test data.

In fact, there are some uncertainties if squares are overlapping too much.
The problem does not require us to find the smallest number of frames, hence it is ok in this case if our solutions produce 2 small squares instead of 1 big square.

My first message was actually right, but it no longer applies - they rejudged the problem apparently, and my program that initially got AC now fails on test 6.

So just skip this thread.

Incidentally, I don't know how to fix my submission yet, but that's another thing :)