Make the physically observed projection match the intended appearance despite wall color, projector gamma, ambient light, blur, occlusion and surface albedo.
- Per-channel monotonic inverse LUT baseline.
- Radiometry dataset capture tooling.
- Learned appearance-compensation MLP scaffold.
- Closed-loop optimization scaffold.
- Prototype: inverse gamma/LUT.
- Usable: measured wall/projector dataset and LUT improvement.
- Polished: spatially varying correction, ambient adaptation, defocus/occlusion masks.
- Advanced: learned forward/inverse display model with online camera refinement.
- Ridiculous: perceptual/semantic closed-loop optimization jointly aware of room geometry, projector limits and generative target.
- Capture real grayscale/RGB/random-patch dataset.
- Compare direct projection vs LUT vs learned inverse on held-out colors/patches.
- Add spatially varying response map and camera exposure lock.
- Add saturation/gamut constraints and shadow/occlusion masks.
- Add defocus compensation baseline.
- Train forward surrogate
C(P, surface, illumination)and validate predictions. - Add online refinement and perceptual objectives (LPIPS/DINO/segmentation) only after RGB baseline.
Observed RGB/Lab error, perceptual error, spatial error maps, saturation rate, convergence time, temporal stability, latency overhead.