Learning to read the style of older research papers and to pull applications out of them is very useful though. It's surprising how often something seemingly overacademized and useless, when translated into code, turns out to be very useful and smart.
I guess I should clarify that I am not afraid of the math heavy papers (image processing is by nature math heavy) but what I meant by overly academic is proving something just to write a paper. For instance sometimes they write a paper on a solved problem to take the overall asymptotic running time down a bit but increase the constant factor and code complexity by a large amount. Think quickSelect vs Deterministic select. Quickselect only breaks down if every random pick of the pivot is the worst one whereas the deterministic select guarantees O(n) but with a constant in the 10's (I could be off on this number but it's high)
Also, they sort conference papers using different criteria (most featured, views, most impact, etc) and hence can quickly find good papers to read or implement.
while the fundamental computer vision algorithms can be found in this free book "Introduction to Computer Vision" by Prof. Mubarak Shah, one of the pioneers in this area.
Both books are available on Sciweavers.