Learning with Manifolds in Computer Vision

Guest Editors

Mohamed Daoudi, IMT Lille Douai, CRIStAL, France

Mehrtash Harandi, Monash University, Australia

Vittorio Murino, University of Verona, Verona, Italy, and Huawei Technologies Ltd., Ireland Research Center, Dublin, Ireland

Paper submission due: January 31st, 2021
First Notification: April 31st, 2021
Revision: Final Decision: August 31st, 2021 
Publication: 2021 (tentative)

Aim and Scope

Manifold Learning (ML) has been the subject of intensive study over the past two decades in the computer vision and machine learning communities. Originally, manifold learning techniques aim to identify the underlying structure (usually low-dimensional) of data from a set of, typically high-dimensional, observations. The recent advances in deep learning make one wonder whether data-driven learning techniques can benefit from the theoretical findings from ML studies. This innocent looking question becomes more important if we note that deep learning techniques are notorious for being data-hungry and (mostly) supervised. On the contrary, many ML techniques unravel data structures without much supervision. This special issue aims at raising the question of how classical ML techniques can help deep learning and vice versa, and targets works and studies investigating how to bridge the gap.

Besides, the use of Riemannian geometry in tackling/modelling various problems in computer vision has seen a surge of interest recently. The benefits of geometrical thinking can be understood by noting that in many applications, data naturally lies on smooth manifolds, hence distances and similarity measures computed by considering the geometry of the space naturally result in better and more accurate modelling. Various studies demonstrate the benefits of geometrical techniques in analysing images and videos such as face recognition, activity classification, object detection and classification, and structure from motion to name a few.

This special issue addresses challenges and future directions related to the application of non-linear manifold and machine learning in computer vision.

Topics and Guidelines

This special issue targets researchers and practitioners from both industry and academia to provide a forum in which to publish recent state-of-the-art achievements in Non-Euclidean geometry and machine learning for computer vision. Topics of interest include, but are not limited to:

● Theoretical Advances related to manifold learning

● Dimensionality Reduction (e.g., Locally Linear Embedding, Laplacian Eigenmaps)

● Clustering

● Kernel methods

● Metric Learning

● Time series on non-linear manifolds

● Transfer learning on non-linear manifolds

● Generative Models on non-linear manifolds

● Subspace Methods

● Advanced Optimization Techniques (constrained and non-convex optimization techniques on non-linear manifolds)

● Mathematical Models for learning sequences

● Mathematical Models for learning Shapes

● Deep learning and non-linear manifolds

● Low-rank factorization methods

● Graph-based Analysis

● Learning via Hyperbolic geometry

And related applications in computer vision (a non-exhaustive list in provided below):

● Face recognition

● Image/video analysis and classification

● Action/activity recognition

● Behavior analysis

● Facial expressions recognition

● Person Re-Identification

● Face generation

● Facial expression generation

● Fine-grained recognition

● Visual inspection