This paper proposes a random coded sampling method for capturing high-speed HDR videos. At each frame, we sample a subset of pixels with predefined random exposure times; other pixels that are not sampled at this frame continue to expose. For example, at frame a, pixels 1, 6, and 9 are read; at frame b (7 frames after frame a), pixels 3 and 7 are read. Color bars indicate exposure length. Our algorithm reconstructs full-frame high-speed HDR videos.
Our exposure scheme has 4 advanatages:
We compare our sampling scheme with conventional global constant exposure settings that operate at a frame rate of 1/4 the high-speed frame rate; the factor 1/4 is used because with current parameters, our random sampling scheme has a compression ratio of 4:1. In particular, we use two constant exposure settings: a long exposure time that corresponds to 4 high-speed frames (referred to as 4x exposure in the following examples) and a short exposure time that corresponds to 1 high-speed frame (referred to as 1x exposure).
For further comparison of our results to the constant 1x exposure videos, we show denoising results of the 1x exposure videos using the CBM3D method. We did not use the video version of this method (CVBM3D) because its output is quantized, leading to poor results in dark regions after applying a curve adjustment.