Incentives in academia as things are is ... uh. Not so awesome.

My expectation from experience when implementing something from a DSP paper is that the result will be unreproducable without contacting the authors for some undisclosed table of magic constants. After obtaining it, the results may match but only for the test images they reported results on. Results on anything else will be much worse.

Also it's normal for techniques from the literature to have computational/memory bandwidth costs two orders of magnitude greater than justified for even their (usually exaggerated) stated levels of performance.

And then their comparison points are almost always inevitably implemented so naively as to make the comparison useless.

It's always difficult because improvements in this domain (like many other engineering domains) are significantly about tradeoffs ... and tradeoffs are difficult to weigh in a pure research environment without the context of concrete applications. They're also difficult to weigh with implementation cleverness having such a big impact particularly since industry heavily drains academia of naturally skilled software engineers.

And as other comments have pointed out, academia is in some sense among the worst of the patent abusers. They'll often develop technology just far enough to lay patent mines around the field, but not far enough to produce something useful out of it. The risk that you spend the significant effort to turn a concept into something usable only to have some patent holder show up with a decade old patent to shake you down is a big incentive against investment.