A good guide and peer group is of utmost importance in tackling problems in any research field. Gain some exposure through a short term research internship at a University/Lab with a good CV group. Learning from the right guys strengthens your fundamentals, shows you where you can anticipate bugs and provides you with the right approach to fix them.

Stay in trend: Follow tier 1 Computer Vision conferences relentlessly: CVPR, ICCV and ECCV. Simply reading the publication is hardly enough. In order to understand the work in context, you must either attend the conference or go through the videos from the conference that are made available online: Youtube videos and Lectures from ECCV 2014.

Beat to death: Discuss the papers vigorously with your peers and find aspects that can be improved. These can be found for a slightly older paper through papers that cite it. Talk to your guide for high level input.

Make stuff work: Get hold of the code for the paper. Authors generally do make code available or are usually willing to share snippets that are crucial to the work. Simply understanding their implementations goes a long way into clarifying several nitty gritty aspects of the system. If you do have to implement the code yourself, fret not. There are plenty of forums: OpenCV Q&A Forum, Stack Overflow and MATLAB Central where you can work out the implementation details. Of course, it's great if you can find a good team with enough experience to work with.

Keep in touch with the theory (Coursera) and you're good to roll.