透過無參考圖像質量評估揀選神經輻射場不確定區域之主動式神經輻射場拍攝

Abstract

作為體積神經運算中搏得驚豔的渲染結果之方法，神經輻射場（NeRF）已在電腦視覺以及圖學上發展迅速。然而，為了得到擬真的畫面，伴隨而來的是較多的訓練資料以及龐大的運算時間，近期地研究方向著重快速的收斂來減少運算時間以及使用額外的限制來達到較少的訓練資料。其中，鮮少有人研究如何從NeRF中省去繁冗的圖像拍攝以及揀選，從而達到逐步構建出質量較好的結果。 在本論文中，我們展示了一個循序揀選最佳影像的方式，來為廣大的視角挑選做了一個建議。為了達到此做法，我們定義了由NeRF產生出來的瑕疵圖片風格，並蒐集由該圖片以及相對應的品質分數等資料，透過無參考圖像質量評估的方式，進而學習了一個模型去判斷NeRF所預測出來圖像的不確定性評比，對於一個選定的場景，我們建立一個相機攝影場域範本，在其上透過此評比協助我們能夠循序的找到重建NeRF場景中必要的視角，重而得到較佳的結果。透過我們方法中推薦的視角選取，我們展示了該方法能夠給予使用者一個拍攝的方向，比起無標準隨機拍攝可以得到更佳的NeRF。

As a fascinating representation of neural volume, Neural Radiance Fields (NeRF) have got significant results for learning to represent 3D scenes. However, to produce photorealistic images, which are accompanied by lengthy training time and more training data, is often an improvement direction like training with fewer datasets or fast convergence to reduce the cost of training time. Besides, capturing real-world scenes requires tons of shooting images from different directions. To reduce the labor burden and clarify the missing part of NeRF, we present a sequence scanning technique that reduces the number of images required, which suggests the view you should take based on what you have already shot. This goal is achieved by scanning at strategically selected Next-Best-Views (NBVs) to capture the object's geometric details progressively. The key of our method is the uncertainty analysis according to the RGB images predicted by NeRF model. We can sequentially capture the missing view through the uncertainty analysis and get a better reconstruction.