Fig. 1. Eight possibilities for the location of the central pixel

Fig. 2. Pseudocode of sub-window box filter algorithm

Fig. 3. HardNet network structure

Fig. 4. Construction process of triplet training set

Fig. 5. Algorithm process

Fig. 6. Construct the scale space

Fig. 7. Pseudo code of circular non-maximum suppression algorithm

Fig. 8. Attention-HardNet

Fig. 9. Experimental reference images (left) and images to be matched (right). (a) Bike image; (b) UBC image; (c) Wall image; (d) Leuven image; (e) Bark image; (f) Boat image; (g) Graffito image; (h) Trees image

Fig. 10. Matching accuracy. (a) Bike image; (b) UBC image; (c) Wall image; (d) Leuven image; (e) Bark image; (f) Boat image; (g) Graffito image; (h) Trees image

Fig. 11. Matching renderings. (a) Bike image; (b) UBC image; (c) Wall image; (d) Leuven image; (e) Bark image; (f) Boat image; (g) Graffito image; (h) Trees image