Block-Based Error Measure for Object Segmentation

Object segmentation is an important research topic in computer vision. The assessment of the quality of the segmentation results is of crucial importance. The conventional performance measure for object segmentation is mainly based on the pixel error between the segmented object and ground truth. The pixel-based error measure does not consider the spatial distribution of segmentation errors, which is essential in semantic processing. This paper presents a novel block-based error measure for evaluating the performance of object segmentation. We first analyze the spatial distribution of segmentation errors and classify them into scattered error and region error. Then we develop a block-based error measure that enhances the contribution of the region error. The mathematical analysis of both error measures is also presented to demonstrate the advantages of the proposed block-based measure.

關鍵字

block-based error measure ； object segmentation ； pixel-based error measure ； relative evaluation

並列摘要

物件分割是電腦視覺中重要的研究主題，而如何評估分割結果的品質是非常重要的議題。傳統的物件分割效能的評估主要都是建立在分割物件與標準物件之畫素錯誤上。這種以畫素錯誤為基礎的量測技術沒有考慮分割錯誤的空間分佈（spatial distribution），而此空間分佈在影像意涵處理上是很重要的。本文提出一用於評估物件分割效能之區塊錯誤為基礎的量測新技術。首先，本文分析分割錯誤的空間分佈，然後將其分類為分散型錯誤與集中型錯誤，最後提出區塊錯誤為基礎的量測技術，此技術強化集中型錯誤的貢獻度。我們也以數學分析顯示此新量測技術的優異性。

並列關鍵字

區塊錯誤量測；物件分割；畫素錯誤量測；相對評估

主題瀏覽