The proposed hybrid "intensity+event" imaging scheme combines a dual-path optical setup to record coded intensity measurements and intermediate event signals simultaneously, and a dual-branch Transformer network to reconstruct dense video frames at high fidelity, achieving 10,000 frames per second.
We present a novel imaging and neural rendering technique that synthesizes videos of light propagating through a scene from arbitrary, dynamic camera viewpoints.
The exit pupil of the main lens plays a crucial role in accurately relating the light field within a standard plenoptic camera to the 3D scene in front of the camera. Ignoring the exit pupil can lead to significant errors in refocusing and depth estimation.
Proposing the Exhaustive Correlation Transformer (ECT) to model unified spatial-spectral correlation and linear dependence for spectral super-resolution.
Novel QDMR framework improves real-world CAC by adapting to target domain with unpaired data.