I'm developing a software with OpenCv Library , my software is based on a paper about the multispectral stereo odometry. After edge extraction, from the image , i have to compute for each extracted feature the edge histogram descriptor. From the paper:
"The spatial component of the descriptor is obtained as follows:
• Select a region of P × P pixels centered at the keypoint of interest from the edge map.
• Divide the region into 16 (4 × 4) subregions.
• Compute local edge histograms for each subregion where five bins are used to categorize an edge: horizontal, vertical, 45◦ diagonal, 135◦ diagonal, and isotropic (no orientation).
The resulting histogram vector formed of 80 bins (4 × 4 × 5) is then normalized. "
My question is: if the keypoint is on the border of the edgeMap, my window (PxP) is out of bound for every P>=3 . Some paper reccomended to use P=100. The same problem occurs also if the keypoint is near to the border.
How should i proceed?