Like others before me HOG Trilinear Interpolation of Histogram Bins I am trying to understand and implement trilinear interpolation for HoG. Dalal's Thesis' here:
http://lear.inrialpes.fr/people/dalal/NavneetDalalThesis.pdf
The issue I am struggling to understand is that as I understand it, the task is of value re-destribution into bins rather than interpolation of known bin values. I am clear about the one dimensional case in the orientation dimension;
for example if I have a known pixel gradient orientation value of 75
and nearest bins of z1=60
and z2=80
(bin step size b=20
), then the pixel value (gradient magnitude in this case, say w=16
) will be distributed at a ratio of 1:3
into the adjacent bins (4 into the 60
bin and 12
into the 80
bin) right?
When it comes to the three dimensional case that requires spatial interpolation of the magnitude, I'm stumped. I fail to understand how the magnitude is redestributed to the appropriate bins. Any clear explanation for someone, who is admitedly slow, will be greatly appreciated.