A fully automated greedy square jigsaw puzzle solver Supplementary material

Figs. 1-4 shows our solver results for for all 20 images used by Cho et al. [1], where each puzzle consists of 432 patches of 28 × 28 pixels. In addition to the final result, these figures also show the results of the initial placement by the placer, the initial segmentation by the segmenter, and the final result after the shifter. The average performance metrics accuracy results for all 20 images is 91% accuracy under the Direct comparison, and 94% under the Neighbor comparison, Importantly, we were able to obtain 100% accurate solutions for 65% of the images. Figs. 5-8 shows our solver results for 20 images from McGill image database [2], where each puzzle consists of 540 patches of 28×28 pixels. In addition to the final result, these figures also show the results of the initial placement by the placer, the initial segmentation by the segmenter, and the final result after the shifter. The average reconstruction accuracy for the direct and neighbor comparison was 83% and 91% respectively, while 50% of the images were constructed with 100% accuracy. Figs. 9-12 shows our solver results for 20 images from a self created database, where each puzzle consists of 805 patches of 28 × 28 pixels. In addition to the final result, these figures also show the results of the initial placement by the placer, the initial segmentation by the segmenter, and the final result after the shifter. The average reconstruction accuracy for the direct and neighbor comparison was 80% and 90% respectively, while 35% of the images were constructed with 100% accuracy. Each one of Figs. 13-14 show our solver results for 3 images, where each puzzle consists of 2360 and 3300 parts of 28× 28 pixels each, respectively. The jigsaw solver MATLAB code and images are available for the use of the community [3].