人工特徵與深度學習特徵在基礎矩陣估算上之效能表現

Abstract

特徵提取在攝影機定位領域有廣泛的應用，特徵提取的精確度對於追蹤攝影機至關重要。為此，本研究透過兩視點跟蹤流估計基矩陣的比較，驗證了特徵之性能。在這個過程中，我們提出了一個方法來比較特徵運用在整個領域上的性能，並使用一個新的概念，規範對稱幾何距離曲線。除了常用的數據集，包括視點或亮度的變化，我們還使用專門用於低曝光條件下的數據集來評估相機傳感器的靈敏度。低曝光和噪聲的方法是用來定義強健特徵的另一種視角，對於未來諸如自動駕駛汽車或增強現實等應用中的特徵提取方法的發展有重大影響。

Feature-based localization method is widely used in the camera-based localization field, and accurate feature extraction is crucial for obtaining the precise camera tracking result. To this end, in this study, the performance of the feature was verified through a comparison of the estimated fundamental matrices following a two-view tracking flow. In this paper, we present a new evaluation metric, the NSGD error curve, to compare the overall performances of various feature methods. In addition to use the comparative evaluation dataset for viewpoint or luminance changes, we also applied an low exposure and noise specific dataset to evaluate the performances of the feature extraction methods in low luminance situation. This new approach to low exposure and noise, another criteria that defines robust features, will have significant meanings for researching feature methods in applications such as Autonomous Vehicles or Argumented Reality.