Robust image stitching with multiple registrations ECCV2018 本文使用多個 registrations 來增強影象拼接的效果

標準的影象拼接流程一般為：1）得到一個 single registration（這裡我們使用 registration for an arbitrary (potentially non-rigid) image transformation, and homography for a line-preserving image transformation.）；2）經過 image warp，將影象 對映到同一個座標系下面 進行 seam finding step；3） 使用 blending procedure 消除一些 unpleasant artifacts， 例如 minor misalignment, or differences in color or brightness due to different exposure or other camera characteristics

在單個registration中難以解決 parallax or motion問題，而後續步驟 seam finding 也不能補償上述問題引入的誤差

本文提出使用多個 registration，The seam finding stage is then free to choose different registrations for different regions of the composite output image. Note that as any registration can serve as a candidate under our scheme, it represents a generalization of methods that attempt to find a single good registration for stitching

這裡我們沒有使用傳統的 seam finding approach，因為給定多個 registration 時，傳統的 seam finding approach 不能很好的工作。不工作的原因分析如下： 1）traditional seam finding treats each pixel from the warped image equally，但是在我們 multiple registration 演算法中，each of them only provides a good alignment for a particular region in the image，所以我們需要考慮 pixel-level alignment quality in the seam finding phase 2） seam finding 有時會失敗導致物體重影，對於 multiple registrations 則對應物體出現多次

1.2 Problem formulation and our approach 問題的描述以及解決方法 這裡我們採用常用的影象拼接表示方式，sometimes called perspective stitching or a flat panorama ，將影象 1 作為基準影象，然後將影象2變換到基準影象座標系下面，再講內容疊加到影象1上面

這裡我們不是隻送一個 warped ω(I1 ) 到後續步驟 seam finding phase 裡，我們提取一組 多個 warping ω 1 (I1 ),…,ω N (I1 )，其中每個 warping ωi(I1) 只負責對齊 兩個影象中的某一個區域。然後 Then we will formalize a multi-label MRF problem for seam finding，We will get the optimal seam by minimizing the energy function E(x) ，最後我們 adopt Poisson blending [9] to smooth transitions over stitching boundaries when generating the final result.

我們如果使用 traditional MRF energy stitch multiple registrations 得到不是很好得結果。we propose the improved MRF energy by adding (1) a new data term that describes our confidence between different warping proposals at pixel p and (2) several new smoothness terms which attempt to prevent duplication or tearing.

本文提出的演算法速度應該不會太快的！

11