210 likes | 363 Views
Presentation 4 Ruben Villegas 06/12/2012. Approximating crowd density taking advantage of occlusion. Papers read.
E N D
Papers read • Density-aware person detection and tracking in crowds by MikelRodriguez, Ivan Laptev, Josef Sivic, Jean-Yves Audibert, EcoleNormaleSuperieure, INRIA Imagine, LIGM, UniversiteParis-Est • Part-based Multiple-Person Tracking with Partial Occlusion Handling
Score map values • The intense red in the score map represents high detection scores returned by the part-base model in the algorithm • The blue represents low detection scores. • The green and yellow are assumed to be only parts of humans but not the entire body, since the scores sit in the middle
First try to get the occlusion map • score.*inv_score • The assumption was that this will multiply high values with high values inverted and same for low values, therefore the middle values will stay. • Since inv_score is (score)^(-1) before being converted to range between 0 nad 1, if the max value in score was 5, inv_score = 0.2, and in score converted between 0 and 1, 5 will be 1. Therefore, 5 will map to 1 and the middle values will be ie. 0.3*0.5 = 0.18 and 1*0.25 = 0.25 we do not accomplish our goal of making middle values higher.
Process to get middle scores • Get the max and min values from score. • Go through every element in score and subtract from max and min separately. • Compare results and pick the lowest one. • Save all those results in a new matrix.
Getting approximate occlusion map • To get the occlusion map, we want to make sure the partially detected people are pronounced. • Adding the score map and the middle values map will give us an approximate occlusion map • To check the result it is used as a mask on the original picture keeping the values that are higher than the average of the max and min values in the occlusion map.
What’s wrong with the occlusion map? • For every true detection(high scores), there are many middle values around them which refer to the same people,and they shouldn’t be considered in the occlusion map. • To solve this we will incorporate the scores from each part separately in the occlusion map.
Incorporating the parts • Every pixel in the score map represents the sum of the score of all the parts detection scores. • Incorporating the score for each part directly into the occlusion map will give us a better occlusion map by emphasizing each part in it.
All parts score map added up Score map from the entire Model detection