Dense SIFT for ghost-free multi-exposure fusion

Two new dynamic fusion examples are shown in Figs. 1-2. In Fig. 1, the input LDR image sequence consists of four exposures, in which both the first and the second contain a walking people but with different locations. Fig.1(b) and (c) shows the fusion results of the image gradient (IG) and the proposed DSIFT based methods, respectively. In this materia, only the “weighted-average” strategy is applied since it usually has better deghosting performance than the “winner-take-all” strategy. It can be seen that the DSIFT method clearly have a better performance than the IG method in terms of ghost removal. In Fig. 2, the input image sequence has five exposures. Both the first and last LDR images contain a walking people almost with the same location. There is also a camera backpack in the second source images. The fusion results of the IG and DSIFT methods are shown in Fig. 2(b) and (c), respectively. We can see that the proposed method successfully removes the ghost artifacts and outperforms the IG method clearly. In general, our method can obtain a satisfactory performance with about four LDR images when the moving objects are not largely overlapped.