Square jigsaw puzzle solver based on [1]. Generates solutions for the Huawei Honorcup Marathon 2
- Python 3.6+
- Create a new virtual environment and install dependencies:
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
- Dataset can be downloaded from https://yadi.sk/d/BmnFhkaD1Vy4vA
- Extract the data into
data/
in the project root - To run the solver on an image cut into 8x8 parts and scrambled:
$ python main.py data/data_train/64/1623.png 8 8
1623.png
50 8 43 45 25 57 52 35 11 36 56 63 40 41 5 12 0 21 1 27 13 46 33 17 31 34 58 60 15 59 14 38 54 23 26 49 18 37 55 16 32 42 47 6 3 53 7 2 62 61 30 39 29 44 48 24 22 9 10 28 20 51 19 4
1623.png,5.762885261,0.9553571428571429,1
The first two lines of the output are printed to stdout and are in the contest
submission format, while the 3rd line was actually printed to stderr and
contains some statistics about the solving process in the format basename,time taken,score,iterations
.
- To view the unscrambled image, you can pass the tile indices (second line of
the output above) to
rearrange_image.py
via standard input. The rest of the arguments are the input image, number of rows and columns and output file.
$ pbpaste | python rearrange_image.py data/data_train/64/1623.png 8 8 out.png
I used GNU Parallel to run the solver on the 300-600 images in each group. For
example, to generate solutions for all the images in the
data/data_test2_blank/64
folder:
mkdir -p results/A3
ls data/data_test2_blank/64 | parallel "python main.py data/data_test2_blank/64/{} 8 8 > results/A3/{}.txt"
lhis writes the solutions for each image into the results/A3
folder while
printing statistics to the terminal through standard error.
To generate the final submission, just concatenate all the results together:
cat results/A3/* > results/A3.txt
[1]: D. Pomeranz, M. Shemesh, and O. Ben-Shahar, A fully automated greedy square jigsaw puzzle solver, In the Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), Colorado Springs, USA, June, pp. 9-16, 2011. See also the project page for more demos, info, and code