A photograph gives each pixel one color, but a lens receives light from many directions. The image records the result of combining that light under a particular focus setting. A light-field representation keeps some of the directional information separate. That extra dimension is what makes the idea so interesting for both photography and VR.
The light-field projects I collected point toward a different way of thinking about an image. Instead of asking only what color belongs at a position, ask which ray passes through that position and where it is heading. A useful simplified model describes a ray by its intersection with two planes. Two coordinates locate it on one plane, and two locate it on the other.
This sounds abstract until you imagine a row of cameras looking at the same scene. A nearby object changes position substantially from one camera to the next. A distant object changes less. The images contain both appearance and a record of how that appearance shifts across viewpoints.
To refocus, choose a depth and shift the camera images so that points at that depth line up. Average the aligned images. The selected surface reinforces itself because its samples agree. Points at other depths remain misaligned and spread out. In this simplified construction, focus is a choice made during reconstruction rather than a decision permanently baked into one exposure.
The amount of data matters. Two cameras do not magically describe every view between and around them. Surfaces can appear in one camera while being hidden in the other. Thin structures may move several pixels between samples. A reconstruction must interpolate where it has evidence and cope with gaps where it doesn't.
There is a useful tradeoff between spatial and angular sampling. Suppose the sensor budget is fixed. Recording more directions can mean spending fewer samples on the detail within each view. The result may offer a more interesting family of viewpoints while giving each viewpoint less resolution than a conventional photograph. Calling it a higher-quality image misses the actual exchange.
For VR, even a small region of valid movement can change the experience. When I lean slightly, nearby objects should move differently from distant ones. A flat panorama supports looking around from one fixed viewpoint, but it does not contain the observations needed for that translation. A sufficiently sampled light field can support some movement without requiring a complete conventional mesh of the scene.
The word some is doing real work. Move outside the captured region and the representation has to extrapolate. Look behind a surface that every camera saw only from the front and there may be no evidence at all. Light fields expand the set of questions an image can answer. They don't remove the limits imposed by where the cameras were placed.
I find refocusing the clearest demonstration because it exposes the representation directly. The blur is no longer merely an effect painted over a picture. It emerges from how rays are grouped, aligned and combined. The image contains several possible photographs, and reconstruction chooses which one to reveal.