How does it compare to dlib [1], ccv [2] etc.?
ccv is moved on to be all-in on neural-network based computer vision and all development effort happened in that area. (Similar can be said for dlib but at least the author still implement random black-box optimizers from time-to-time).
SOD seems try to reimplement some of OpenCV but TBH, these are not "modern" at all. These are 15 year-old algos while works, the useful domain is shrinking and people who interested in these also might be looking for hardware-accelerated options (AVX-512 / CUDA etc). Some very specific classic heavy CV applications (such as SLAM) are quickly moving to their own dedicated libs (such as Cartographer from Google or ORB-SLAM).