algorithm - Viola Jones threshold value Haar features error value -

i have read viola paper 2004. in 3.1 explain threshold calculation. super confused. reads

for each feature, examples sorted based on feature value

question1) sorted list list of haar feature values calculated integral image of examples. if have feature , 10 images(positive , negative). 10 results associated each input image.

the adaboost optimal threshold feature can computed in single pass on sorted list. each element in sorted list, 4 sums maintained , evaluated: total sum of positive example weights t +, total sum of negative example weights t −, sum of positive weights below current example s+ , sum of negative weights below current example s−

question 2) purpose of sorting. guess 1 highest 1 describes image best. algorithmically how affect (s- s+ t+ t-).

question3) sorted list calculate (s- s+ t+ t-). mean each entry holds own (s- s+ t- t+) or there 1 (s- s+ t- t+) whole list.

the error threshold splits range between current , previous example in sorted list is: e = min ( s+ + (t − − s−), s− + (t + − s+)) ,

question4) answers previous question not sure. in order have "e "for each input image. need maintain (s- s+ t- t+) each entry in list. "e" after calculate n of them (one each image) feature.

thanks in advance, please let me know if confusing or need more clarification questions.

question1) sorted list list of haar feature values calculated integral image of examples. if have feature , 10 images(positive , negative). 10 results associated each input image.

you 10 results feature, 1 result associated each input image. each image marked positive or negative.

question 2) purpose of sorting. guess 1 highest 1 describes image best. algorithmically how affect (s- s+ t+ t-).

the image highest 1 highest response feature. sort based on response, not based on weight.

the reason sort them 2 of things trying calculate "the sum of positive weights below current example s+ , sum of negative weights below current example s−". if list sorted can keep running sum, , @ each point, examples weights have added sum until have feature response less (i.e. "below") current example. doesn't work if list isn't sorted. can evaluate error associated using response level halfway between example , next 1 threshold.

question3) sorted list calculate (s- s+ t+ t-). mean each entry holds own (s- s+ t- t+) or there 1 (s- s+ t- t+) whole list.

there 1 s- , 1 s+ per example, because it's "the sum of positive weights below current example". t+ , t- calculated whole list, don't know why need maintain each element.

question4) answers previous question not sure. in order have "e "for each input image. need maintain (s- s+ t- t+) each entry in list. "e" after calculate n of them (one each image) feature.

you chose minimum out of of them, , that's optimal place put threshold (which midpoint of responses of 2 examples), because has minimum error (false positives + false negatives). btw, reason there 2 choices @ each point (i.e. e = min ( s+ + (t − − s−), s− + (t + − s+)) ) can choose whether make threshold values above response level considered positive (the first term), or values below considered positive.

if it's former, s+ false positives, (t- - s-) false negatives. if it's latter, s- false negatives , (t+ - s+) false positives.