A complex scene contains far more information than any one viewer can see. The back of a wall, the underside of a distant object and millions of tiny details may never contribute to the final image. A renderer normally carries a general-purpose scene because it doesn't know in advance where the camera will go.
Seurat asks what happens when that freedom is limited. Its input is a collection of color and depth observations of a scene from viewpoints within a defined region. It produces a simplified representation intended to preserve the appearance seen from that region while rendering efficiently on mobile VR hardware.
The viewing region, often described as a headbox, is a particularly useful constraint. It allows the viewer to lean and move locally, while avoiding the much harder requirement that every possible position in the original scene remain valid. The output only needs to answer a restricted family of visual questions.
Imagine a museum scene viewed from a small area in front of a display. The viewer can move their head enough to see parallax around the objects. They cannot walk behind the walls or inspect the underside of every shelf. A representation designed for that experience can devote its resources to surfaces and appearances visible from the allowed area.
This is different from reducing every object's polygon count independently. An object-centered simplifier tries to preserve a model's shape across many uses. A view-dependent scene representation can make more aggressive decisions because it knows which views matter. It may keep an image feature that is visually important while discarding geometry whose contribution is invisible.
The captured appearance also contains lighting and shading. That helps the runtime avoid recomputing expensive effects, but it introduces a constraint: changing the scene's illumination isn't automatically equivalent to relighting the original materials. A representation built from observations carries the conditions under which those observations were made.
For an interactive experience, I would separate the captured environment from elements that must respond dynamically. A static architectural setting is one candidate. A moving character, an object the user can pick up, or a light the user can switch needs additional treatment. The boundary should follow the intended interaction, not merely the easiest export path.
The headbox itself becomes part of level design. If the experience encourages the viewer to step beyond it, the content and the representation disagree. A scene can be visually excellent inside its valid region and fail abruptly outside it. Good staging can make the permitted movement feel natural rather than like an arbitrary technical fence.
I like Seurat as an example of optimizing the question before optimizing the answer. Instead of asking a mobile device to reproduce an unrestricted production scene, define the views the experience actually needs. The resulting freedom is smaller, but within that space the visual budget can go much further.