Abstract
This article has presented a novel method for detecting mesh saliency, a measure that tries to be in accord with human perception. Unlike previous methods which operate in the spatial domain, we capture information corresponding to mesh saliency in the frequency domain. Previous methods relying on center-surround local operators tend to capture local saliency, while the proposed spectral mesh saliency method outputs a saliency map which captures globally salient, primary features. We have also considered the underlying reasons why spectral mesh analysis is useful for saliency detection. We have demonstrated how incorporating the proposed mesh saliency can both visually and quantitatively improve the results of several graphics applications such as mesh simplification, mesh segmentation, and scan integration. Note that we do not claim that our saliency measure is superior to other measures such as mesh curvature or shape index in all respects-just that spectral mesh saliency is another useful tool.We hope our work will inspire further investigation into frequency-based techniques for mesh saliency detection. Future work will focus on how to incorporate high-level cues into saliency detection. This is motivated by the limitations of the current framework. As shown in Figure 26, we fail to detect the breasts as salient in a woman model. A challenge for combining such highlevel cues of human perception with a low-level computational model is that some of them are inconstant, typically depending on the gender and the age of people involved in the user study. For the glasses, our method does not capture the central points of interest in each lens suggested by Chen et al. [2012]. Ultimately, however, such high-level cues cannot be captured by a geometric approach alone, as they rely on semantic information. Human users expect to see eyes at the centers of the lenses, as eye contact is an important medium of communication between humans; breasts provide an easy clue to help distinguish whether a stranger is male or female. We are preparing some psychological experiments to further investigate the striking similarities of mesh saliency computed via spectral processing and human-provided results. It will also be interesting to explore further novel applications of mesh saliency such as range image registration and shape from shading.
Original language | English |
---|---|
Article number | a6 |
Journal | ACM Transactions on Graphics |
Volume | 33 |
Issue number | 1 |
Early online date | 31 Jan 2014 |
DOIs | |
Publication status | Published - Jan 2014 |
Keywords
- Applications
- Mesh saliency
- Mesh segmentation
- Mesh simplification
- Spectral mesh processing
- Algorithms
- mesh segmentation
- mesh simplification
- mesh saliency
- spectral mesh processing
Fingerprint
Dive into the research topics of 'Mesh Saliency via Spectral Processing'. Together they form a unique fingerprint.Profiles
-
YONGHUAI LIU
- Computer Science - Professor of Computer Games & Graphics
- Health Research Institute
Person: Research institute member, Academic