People working in computer vision/graphics may have very different backgrounds: computer programming, high-performance computing, signal processing, image analysis, optimization, machine learning, etc.
Novel contributions can be wide: from a very fast GPU/CPU algorithm to render textures (I had a reference in mind which I can't recall now) to semantic segmentation. Origins of ideas may range from huge expertise, trials on billion of images to profound thoughts, through hardware algorithmic wisdom. Your question is far-reaching, and can dig into the depth of the psychology of scientific invention, well-described by some authors in mathematics and physics (I have Raymond Poincaré and Jacques Hadamard in mind). Here are a couple of thoughts.
The ability to (non exclusive):
- be and remain curious: listen a lot, don't pretend too much, don't trust doxa (everybody knows - even yourself - that..., nobody does this), look at novelties in other fields
- test and probe:hardware, data, hypothesis, models... to the limit (add too much noise, down-sample a lot, reduce learning set, cut the power),
- explore unconventional ideas: transfer knowledge from one domain to the other, ask "what if I do..." (I got from Isaac Asimov that most of science is not like Archimedes' "Eureka (I found it)",but often "Ooh, that funny"), learn XXX that strange theory, go to seminars you are not truly interesting in, or above your level
- have people you trust to talk to: especially those who disagree, or dare to say "you're going nowhere", hire good colleagues, better students
- take notes, talk to yourself, don't be discouraged: nice works can take decades
- teach: basic student questions can drive you nuts and shake your confidence
- corollary: have some solid knowledge somewhere (law of excessive learning: you can teach at one level only if you know the level above)
- breathe and sleep: get away from the machine, take some fresh air, wander in woods
- be patient, but remain aware of your progress: one gain confiance progressively. The first "original" ideas I got were already discovered 30 years before. I'm grateful now went I read a 1-year paper that implements what I am thinking of.
- have principles: it is OK to think "This should work!" even if the evidence is shallow. Think about the "AI winter".
Those can be good incentives. On my side, I am lucky to (think that I) have a relatively broad "generic" scientific background (to connect far-fetched domains) and one or two technical skills that I dug deep into. Thus, I am not afraid of switch domains. One main risk I see is siloing: when collaborating with other, it is tempting to (and thus, try to avoid):
- remain in our confidence zone,
- trust others in their domain (as we would like to be treated, with respect).
One of my few scientific prides (not deep, but quite intriguing) was to work with real-time computing people (my programming skills are very limited), and discover that some "dogma" in the domain was false. My contribution was initially very small, on standard extrapolation, which I though I mastered. Some colleagues trusted me and implemented the stuff. But one colleague who did not understand the derivation asked me again, and again, and again, to understand every step that I thought was "evident". So I reduced the formulas down to the bone. With a couple of image compression tricks, we ended up with a scheme that shook the initial dogma in cyber-physical simulation. Not theoretically, but enough to feel we had "something new".