We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.CV

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computer Vision and Pattern Recognition

Title: Geometric Graph Representation with Learnable Graph Structure and Adaptive AU Constraint for Micro-Expression Recognition

Abstract: Micro-expression recognition (MER) is valuable because the involuntary nature of micro-expressions (MEs) can reveal genuine emotions. Most works recognize MEs by taking RGB videos or images as input. In fact, the activated facial regions in ME images are very small and the subtle motion can be easily submerged in the unrelated information. Facial landmarks are a low-dimensional and compact modality, which leads to much lower computational cost and can potentially concentrate more on ME-related features. However, the discriminability of landmarks for MER is not clear. Thus, this paper explores the contribution of facial landmarks and constructs a new framework to efficiently recognize MEs with sole facial landmark information. Specially, we design a separate structure module to separately aggregate the spatial and temporal information in the geometric movement graph based on facial landmarks, and a Geometric Two-Stream Graph Network is constructed to aggregate the low-order geometric information and high-order semantic information of facial landmarks. Furthermore, two core components are proposed to enhance features. Specifically, a semantic adjacency matrix can automatically model the relationship between nodes even long-distance nodes in a self-learning fashion; and an Adaptive Action Unit loss is introduced to guide the learning process such that the learned features are forced to have a synchronized pattern with facial action units. Notably, this work tackles MER only utilizing geometric features, processed based on a graph model, which provides a new idea with much higher efficiency to promote MER. The experimental results demonstrate that the proposed method can achieve competitive or even superior performance with a significantly reduced computational cost, and facial landmarks can significantly contribute to MER and are worth further study for efficient ME analysis.
Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
Cite as: arXiv:2205.00380 [cs.CV]
  (or arXiv:2205.00380v1 [cs.CV] for this version)

Submission history

From: Jinsheng Wei [view email]
[v1] Sun, 1 May 2022 02:20:43 GMT (12587kb,D)

Link back to: arXiv, form interface, contact.