I am currently reading the research paper Image Crowd Counting Using Convolutional Neural Network and Markov Random Field by Kang Han, Wanggen Wan, Haiyan Yao, and Li Hou.
I did not understand the following context properly:

Formally, the Markov random field framework for the crowd counting can be defined as follows (we follow the notation in [18]). Let P be the set of patches in an image and C be a possi- ble set of counts. A counting c assigns a count c p ∈ C to each patch p ∈ P. The quality of a counting is given by an energy function:

E(c) = ∑ D p (c p ) + ∑ p∈P V (c p − c q ) . . . (2) (p,q)∈N

where N are the (undirected) edges in the four-connected image patch graph. D p (c p ) is the cost of assigning count c p to patch p, and is referred to as the data cost. V (c p −c q ) measures the cost of assigning count c p and c q to two neighboring patch, and is normally referred to as the dis- continuity cost. For the problem of smoothing the adjacent patches count, D p (c p ) and V (c p − c q ) can take the form of the following functions: D p (c p ) = λ min((I(p) − c p ) 2 , DATA K) . . . (3) V (c p − c q ) = min((c p − c q ) 2 , DISC K) . . . (4) where λ is a weight of the energy items, I(p) is the ground truth count of the patch p, DATA K and DISC K are the truncating item of D p (c p ) and V (c p − c q ), respectively.

Can anyone explain the above part in detail and give me a detailed insight on how should I implement this part of the project?

  • $\begingroup$ Is there any way in which you could come up with a more descriptive title? Particularly for others who face the same issue as you do. Right now the title is so generic and non-descriptive, that any great answers you may receive will not be found by others in a search. $\endgroup$
    – Bart
    Sep 13, 2018 at 14:20
  • $\begingroup$ @Bart I have edited the title. $\endgroup$
    – Ronith
    Sep 13, 2018 at 14:23


You must log in to answer this question.

Browse other questions tagged .