Paper

Efficient Global 2D-3D Matching for Camera Localization in a Large-Scale 3D Map

Motivation

  • large-scale problem lead to ambiguous matches
    • 3D points can be visually similar or even identical (repeated structure)
    • ambiguous matches are almost inevitable
  • local search will only obtain sub-optimal solution
    • take account of similarities between 2D-3D matches

Proposed scheme

More details

I have some questions about this paper

  • “Covisibility” is the link between each two points or point clouds in the global map?
    • 雖然ref都是定義任兩點之間，但我認為是任兩點雲之間，我的原因如下:
    • 可能點的數量大到無法將Cij寫成矩陣形式？
    • p(t + 1) = αCp(t) + (1 − α)q。這個式子想要做的事情，應該是想辦法把目前的feature points分到某一個global submodels中。(比較符合Introduction提到結構相似產生多個歧義的matching pairs情境。)
    • 因為從點雲變成點，C裡的值之間的差距會變的很小，且每個值應該都會很小。(1 − α)q這項就會dominate，單靠q值得到最後的評分。但q值是跟matching pairs的分布有很大的關係，越多的features或越相近的features連結到這個點，該點q值就會變大。因此這樣又回到問題本身，也就無法解決歧義的問題。