we are looking for simple but effective models, which means even the top conferences' best paper may be rubbish from our sight.
Academic research cares about the performance and novel method, the idea is the most important.
However, industrial environment turn us to be more practical and realistic, we prefer the models with less time consumption (space is enough, don't worry about that) and calculation (no gpu in almost time).
new model > new techniques > new methods(conbines serveral methods and models, this is disgusting).