The registration problem of object motion in overexposed areas.

[lllllialois]

Hello, author. I am a student who has just started learning about multi-exposure fusion. During my studies, I found that significant motion of objects in the overexposed areas of normal exposure frames can easily lead to ghosting in the final HDR image. When using registration methods, due to the large motion, the frames seem unable to align with each other. Additionally, from a global alignment perspective, it is not globally aligned. In light of this situation, are there any good solutions available?

[ltkong218]

This is a challenge case for multi-exposure hdr fusion, and global alignment can not deal with large local moving areas.

[lllllialois]

Do you think that using end-to-end paired data for the network to learn the registration of overexposed areas by itself, compared to not artificially adding some prior guidance (similar to optical flow fields), would be more effective? Additionally, would direct fusion in overexposed areas yield better results,like the cloud layer in the upper left corner of case 134 in the test set?

[ltkong218]

For the first question, networks use or do not use some prior guidance (like optical flow) are both end-to-end trainable. The effectiveness depends on motion characteristics, data volume and scene diversity. Direct fusion in overexposed areas can get relatively good results if registration is successful. However, local motion makes registration challenging, and refinement is required.