Crowd Counting Based on Multiresolution Density Map and Parallel Dilated Convolution

Figure 6

Experimental results. (a) The original image; (b) the manual annotation image; (c) the high-resolution density map produced by the MDPDNet; (d, e, and f) the three low-resolution density maps with L1, L2, and L3, respectively. We can see that the higher resolution the branch has (L3 > L2 > L1), the more accurate the positioning of the small-scale portrait of the image is.

(a)

(b)

(c)

(d)

(e)

(f)