Image quality assessment for visual object segmentation

An image segmentation model is deemed successful if reflects how a human would segment an image into foreground and background (i.e. be as close to the ground truth (GT) as possible). To train a successful model, the evaluation of binary foreground maps (FM) plays an important role. Recent evaluatio...

全面介紹

Saved in:
書目詳細資料
主要作者: Ng, Darryl Jingheng
其他作者: Lin Weisi
格式: Final Year Project
語言:English
出版: Nanyang Technological University 2024
主題:
在線閱讀:https://hdl.handle.net/10356/175092
標簽: 添加標簽
沒有標簽, 成為第一個標記此記錄!
機構: Nanyang Technological University
語言: English
實物特徵
總結:An image segmentation model is deemed successful if reflects how a human would segment an image into foreground and background (i.e. be as close to the ground truth (GT) as possible). To train a successful model, the evaluation of binary foreground maps (FM) plays an important role. Recent evaluation methods like E-measure fail to consider the semantic meanings of images for binary FM evaluation. In this paper, I investigate the feasibility to perform binary FM evaluation using an attention map (Attention Measure), as well as its effectiveness when integrated into E-measure, to make up for E-measure’s lack of semantic consideration in the form of a new measure, Aligned Attention Measure. Attention Measure showed promising results by itself, while Aligned Attention Measure achieved improvements to E-measure and managed to reflect a better human visual system (HVS) correlation in terms of ranking binary FMs. Future work will explore the fine-tuning of this metric to minimise variability as well as improve its parity with HVS.