Why your error rate is higher than original paper?

[deryaguler95]

I checked your density map. And they are so different from original. Is this reason that error rate is higher the original paper?

[dmburd]

The exact algorithm for density map generation is not present in the original repos. I followed the procedure described in the MCNN paper (CVPR 2016). Values of some of the parameters are not given in the paper, so I have chosen them in such a way that my density maps look more or less similar to the xhp ones. Of course, a different algorithm can be used (see here, for example). It's hard to say exactly why I got higher error rates than reported in the original paper, but it's likely because of the limited training time that was available for me. Longer training time should yield lower error rates.

[deryaguler95]

The epoch is 1000. I think it is not a reason for higher error rate. But I do not know :) May be the main point is density map generation.

[dmburd]

Are you planning to reproduce the density map generation used in the official repo?

[deryaguler95]

Yes I am. I work on these topic. I think main point is density map generation. How did you find optimum parameters for this project?