Abstract: We propose bookworm continual learning(BCL), a flexible setting where unseen classes can be inferred via a semantic model, and the visual model can be updated continually. Thus BCL generalizes both continual learning (CL) and zero-shot learning (ZSL). We also propose the bidirectional imagination (BImag) framework to address BCL where features of both past and future classes are generated. We observe that conditioning the feature generator on attributes can actually harm the continual learning ability, and propose two variants (joint class-attribute conditioning and asymmetric generation) to alleviate this problem.

Abstract: Domain adaptation (DA) aims to transfer the knowledge learned from a source domain to an unlabeled target domain. Some recent works tackle source-free domain adaptation (SFDA) where only a source pre-trained model is available for adaptation to the target domain. However, those methods do not consider keeping source performance which is of high practical value in real world applications. In this paper, we propose a new domain adaptation paradigm called Generalized Source-free Domain Adaptation (G-SFDA), where the learned model needs to perform well on both the target and source domains, with only access to current unlabeled target data during adaptation. First, we propose local structure clustering (LSC), aiming to cluster the target features with its semantically similar neighbors, which successfully adapts the model to the target domain in the absence of source data. Second, we propose sparse domain attention (SDA), it produces a binary domain specific attention to activate different feature channels for different domains, meanwhile the domain attention will be utilized to regularize the gradient during adaptation to keep source information. In the experiments, for target performance our method is on par with or better than existing DA and SFDA methods, specifically it achieves state-of-the-art performance (85.4%) on VisDA, and our method works well for all domains after adapting to single or multiple target domains.

Abstract: Multimodal representations and continual learning are two areas closely related to human intelligence. The former considers the learning of shared representation spaces where information from different modalities can be compared and integrated (we focus on cross-modal retrieval between language and visual representations). The latter studies how to prevent forgetting a previously learned task when learning a new one. While humans excel in these two aspects, deep neural networks are still quite limited. In this paper, we propose a combination of both problems into a continual cross-modal retrieval setting, where we study how the catastrophic interference caused by new tasks impacts the embedding spaces and their cross-modal alignment required for effective retrieval. We propose a general framework that decouples the training, indexing and querying stages. We also identify and study different factors that may lead to forgetting, and propose tools to alleviate it. We found that the indexing stage pays an important role and that simply avoiding reindexing the database with updated embedding networks can lead to significant gains. We evaluated our methods in two image-text retrieval datasets, obtaining significant gains with respect to the fine tuning baseline.

Abstract: The main objective of this thesis is to highlight the salient object in an image or in a video sequence. We address three important—but in our opinion
insufficiently investigated—aspects of saliency detection. Firstly, we start
by extending previous research on saliency which explicitly models the information provided from the context. Then, we show the importance of
explicit context modelling for saliency estimation. Several important works
in saliency are based on the usage of object proposals. However, these methods
focus on the saliency of the object proposal itself and ignore the context.
To introduce context in such saliency approaches, we couple every object
proposal with its direct context. This allows us to evaluate the importance
of the immediate surround (context) for its saliency. We propose several
saliency features which are computed from the context proposals including
features based on omni-directional and horizontal context continuity. Secondly,
we investigate the usage of top-downmethods (high-level semantic
information) for the task of saliency prediction since most computational
methods are bottom-up or only include few semantic classes. We propose
to consider a wider group of object classes. These objects represent important
semantic information which we will exploit in our saliency prediction
approach. Thirdly, we develop a method to detect video saliency by computing
saliency from supervoxels and optical flow. In addition, we apply the
context features developed in this thesis for video saliency detection. The
method combines shape and motion features with our proposed context
features. To summarize, we prove that extending object proposals with their
direct context improves the task of saliency detection in both image and
video data. Also the importance of the semantic information in saliency
estimation is evaluated. Finally, we propose a newmotion feature to detect
saliency in video data. The three proposed novelties are evaluated on standard
saliency benchmark datasets and are shown to improve with respect to
state-of-the-art.

Abstract: This paper considers the problem of recipe-oriented image-ingredient correlation learning with multi-attributes for recipe retrieval and exploration. Existing methods mainly focus on food visual information for recognition while we model visual information, textual content (e.g., ingredients), and attributes (e.g., cuisine and course) together to solve extended recipe-oriented problems, such as multimodal cuisine classification and attribute-enhanced food image retrieval. As a solution, we propose a multimodal multitask deep belief network (M3TDBN) to learn joint image-ingredient representation regularized by different attributes. By grouping ingredients into visible ingredients (which are visible in the food image, e.g., “chicken” and “mushroom”) and nonvisible ingredients (e.g., “salt” and “oil”), M3TDBN is capable of learning both midlevel visual representation between images and visible ingredients and nonvisual representation. Furthermore, in order to utilize different attributes to improve the intermodality correlation, M3TDBN incorporates multitask learning to make different attributes collaborate each other. Based on the proposed M3TDBN, we exploit the derived deep features and the discovered correlations for three extended novel applications: 1) multimodal cuisine classification; 2) attribute-augmented cross-modal recipe image retrieval; and 3) ingredient and attribute inference from food images. The proposed approach is evaluated on the constructed Yummly dataset and the evaluation results have validated the effectiveness of the proposed approach.

Abstract: Depth can complement RGB with useful cues about object volumes and scene layout. However, RGB-D image datasets are still too small for directly training deep convolutional neural networks (CNNs), in contrast to the massive monomodal RGB datasets. Previous works in RGB-D recognition typically combine two separate networks for RGB and depth data, pretrained with a large RGB dataset and then fine tuned to the respective target RGB and depth datasets. These approaches have several limitations: 1) only use low-level filters learned from RGB data, thus not being able to exploit properly depth-specific patterns, and 2) RGB and depth features are only combined at high-levels but rarely at lower-levels. In this paper, we propose a framework that leverages both knowledge acquired from large RGB datasets together with depth-specific cues learned from the limited depth data, obtaining more effective multi-source and multi-modal representations. We propose a multi-modal combination method that selects discriminative combinations of layers from the different source models and target modalities, capturing both high-level properties of the task and intrinsic low-level properties of both modalities.

Abstract: In many computer vision applications, vignetting is an undesirable effect which must be removed in a pre-processing step. Recently, an algorithm for image vignetting correction has been presented by means of a minimization of log-intensity entropy. This method relies on an increase of the entropy of the image when it is affected with vignetting. In this paper, we propose a novel algorithm to reduce image vignetting via a maximization of the fuzzy entropy of the image. Fuzzy entropy quantifies the fuzziness degree of a fuzzy set and its value is also modified by the presence of vignetting. The experimental results show that this novel algorithm outperforms in most cases the algorithm based on the minimization of log-intensity entropy both from the qualitative and the quantitative point of view.