In the past, high dynamic range (HDR) techniques have only been widely adopted for still imaging systems, due to the lack of cost-effective HDR motion capture devices. Without supporting hardware, it is reasonable to consider a software approach. The authors of this paper try to synthesize an HDR video device using off-the-shelf cameras such as the Basler acA2000-50gc, which costs about $1,500. This kind of conventional low dynamic range (LDR) camera can be programmed to periodically alternate between different exposures along its capture timeline. The idea is that instead of having to resort to expensive, specialized devices that can capture information from all exposures at the same time, as required by HDR, the proposed approach can recover the complete information for a particular frame using information from adjacent frames that are captured with different exposures. This sounds good at first, but is problematic upon reflection. Most importantly, how do we build the correspondence between adjacent frames, especially when they are of different exposures? Conventional patch-based or optical flow methods alone will fail for different reasons, so the authors propose combining the two to minimize the artifacts arising from wrong correspondences. The whole idea is expressed in equation (1). Figure 2, the introduction to section 3, and section 3.3 will be useful to readers who want to implement the algorithm.
I consider this type of work highly related to image fusion. While the latter lacks prior information and is often based on pure intensity, this work takes advantage of the time correlation between adjacent frames, which suggests that it might achieve higher accuracy. Nonetheless, the performance requirement for video applications in terms of time and quality is also strict. Although the results reported in this paper are far from applicable for real time, the authors emphasize the potential for improvement with more careful implementations. This field is just getting started, and many existing techniques can directly contribute to it.