Manifold learning using robust Graph Laplacian for interactive image search

Abstract

Interactive image search or relevance feedback is the process which helps a user refining his query and finding difficult target categories. This consists in partially labeling a very small fraction of an image database and iteratively refining a decision rule using both the labeled and unlabeled data. Training of this decision rule is referred to as transductive learning. Our work is an original approach for relevance feedback based on Graph Laplacian. We introduce a new Graph Laplacian which makes it possible to robustly learn the embedding, of the manifold enclosing the dataset, via a diffusion map. Our approach is three-folds: it allows us (i) to integrate all the unlabeled images in the decision process (ii) to robustly capture the topology of the image set and (iii) to perform the search process inside the mani

Active learning in multimedia annotation and retrieval

“…In contrast to supervised learning, SSL leverages a large amount of unlabeled samples based on certain assumptions, such that the obtained models can be more accurate than those achieved by purely supervised methods. Since SSL and active learning both involve unlabeled data, they have been exploited together in many image/video annotation and retrieval works [Wang et al 2007;Zhu et al 2003b;Sahbi et al 2008]. Zhu et al [2003b] proposed an active learning approach based on a graph-based SSL method, in which the reduction of expected risk of labeling each sample can be predicted without retraining classification model.…”

“…Learning Method/ Sample Selection Criteria CBIR [Bao et al 2009] Graph-Based SSL risk reduction CBIR [Dagli et al 2006] SVM uncertainty; diversity CBIR [Geng et al 2008] SVM uncertainty CBIR [Goh et al 2004] SVM uncertainty; relevance CBIR [Gosselin and Cord 2004] Bayes classifier; k-NN; SVM uncertainty; diversity CBIR [He et al 2004] SVM uncertainty CBIR [Hoi and Lyu 2005] Graph-Based SSL risk reduction CBIR [Panda et al 2006b] SVM uncertainty; diversity CBIR [Settles et al 2007] Multiple Instance Logistic Regression uncertainty; expected gradient length CBIR [Tong and Chang 2001] SVM uncertainty Multiclass Image Annotation [Vijayanarasimhan and Grauman 2008] SVM with multiple instance kernel risk reduction (Continued on next page) [Jain and Kapoor 2009] Probabilistic k-NN uncertainty Multiclass Image Annotation [Joshi et al 2009] SVM uncertainty Multiclass Image Annotation [Sahbi et al 2008] Manifold Learning uncertainty; diversity Multiclass Image Annotation [Yang et al 2009] Multiple kernel learning uncertainty Multiclass Image Annotation [Zhang and Chen 2003] Kernel Regression uncertainty; density Multiclass Video Annotation [Yan et al 2003] SVM risk reduction Multilabel Image Annotation [Sychay et al 2002] SVM uncertainty Multilabel Image Annotation Multilabel learning uncertainty Multilabel Video Annotation [Ayache and Quénot 2007] SVM uncertainty; positivity Multilabel Video Annotaton [Chen et al 2005] SVM uncertainty Multilabel Video Annotation [Hua and Qi 2008] Multilabel learning uncertainty Multilabel Video Annotation [Naphade and Smith 2004a] SVM uncertainty Multilabel Video Annotation [Qi et al 2004] SVM uncertainty; density Multilabel Video Annotation [Song et al 2005] Gaussian Mixture Model uncertainty; density Multilabel Video Annotation Graph-Based SSL uncertainty Multilabel Video Annotation [Vendrig et al 2002] Maximum Entropy Classifier relevance Multilabel Video Annotation [Wang et al 2007] Manifold Learning uncertainty; density; diversity; relevance…”

Interactive retrieval of targets for wide area surveillance

“…al. [15] use a graph-Laplacian approach for embedding the data manifold via a diffusion map based on user-feedback. The approach also uses heuristics based on active learning to select informative examples for user-feedback.…”

Embedding Retrieval of Articulated Geometry Models

“…Hence, it differs from other techniques that simply discard useful higher dimensions [4]. Recently, DM has also been applied to image databases as a transductive relevancy feedback tool [30]. There are also supervised learning methods, referred to as Metric Learning, that exaggerate the distances among data so that the metric produces small distances for objects within the same category and large distances for those of different categories [31].…”