Abstract: Multiple Instance Learning (MIL) has become a hot topic and many different algorithms have been proposed in the last years. Despite this fact, there is a lack of comparative studies that shed light into the characteristics of the different methods and their behavior in different scenarios. In this paper we provide such an analysis. We include methods from different families, and pay special attention to vocabulary-based approaches, a new family of methods that has not received much attention in the MIL literature. The empirical comparison includes seven databases from four heterogeneous domains, implementations of eight popular MIL methods, and a study of the behavior under synthetic conditions. Based on this analysis, we show that, with an appropriate implementation, vocabulary-based approaches outperform other MIL methods in most of the cases, showing in general a more consistent performance.

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: Multimodal image fusion allows the combination of information from different modalities, which is useful for tasks such as object detection, edge detection, and tracking, to name a few. Using the fused representation for applications results in better task performance. There are several image fusion approaches, which have been summarized in surveys. However, the existing surveys focus on image fusion approaches where the application on the loop of multimodal image fusion is not considered. On the contrary, this study summarizes deep learning-based multimodal image fusion for computer vision (e.g., object detection) and image processing applications (e.g., semantic segmentation), that is, approaches where the application module leverages the multimodal fusion process to enhance the final result. Firstly, we introduce image fusion and the existing general frameworks for image fusion tasks such as multifocus, multiexposure and multimodal. Then, we describe the multimodal image fusion approaches. Next, we review the state-of-the-art deep learning multimodal image fusion approaches for vision applications. Finally, we conclude our survey with the trends of task-driven multimodal image fusion.

Abstract: Multi-person tracking can be exploited in applications such as driver assistance, surveillance, multimedia and human-robot interaction. With the help of human detectors, particle filters offer a robust method able to filter noisy detections and provide temporal coherence. However, some traditional problems such as occlusions with other targets or the scene, temporal drifting or even the lost targets detection are rarely considered, making the systems performance decrease. Some authors propose to overcome these problems using heuristics not explained
and formalized in the papers, for instance by defining exceptions to the model updating depending on tracks overlapping. In this paper we propose to formalize these events by the use of a state-graph, defining the current state of the track (e.g., potential , tracked, occluded or lost) and the transitions between states in an explicit way. This approach has the advantage of linking track actions such as the online underlying models updating, which gives flexibility to the system. It provides an explicit representation to adapt the multiple parallel trackers depending on the context, i.e., each track can make use of a specific filtering strategy, dynamic model, number of particles, etc. depending on its state. We implement this technique in a single-camera multi-person tracker and test
it in public video sequences.

Abstract: Multi-label zero-shot learning strives to classify images into multiple unseen categories for which no data is available during training. The test samples can additionally contain seen categories in the generalized variant. Existing approaches rely on learning either shared or label-specific attention from the seen classes. Nevertheless, computing reliable attention maps for unseen classes during inference in a multi-label setting is still a challenge. In contrast, state-of-the-art single-label generative adversarial network (GAN) based approaches learn to directly synthesize the class-specific visual features from the corresponding class attribute embeddings. However, synthesizing multi-label features from GANs is still unexplored in the context of zero-shot setting. When multiple objects occur jointly in a single image, a critical question is how to effectively fuse multi-class information. In this work, we introduce different fusion approaches at the attribute-level, feature-level and cross-level (across attribute and feature-levels) for synthesizing multi-label features from their corresponding multi-label class embeddings. To the best of our knowledge, our work is the first to tackle the problem of multi-label feature synthesis in the (generalized) zero-shot setting. Our cross-level fusion-based generative approach outperforms the state-of-the-art on three zero-shot benchmarks: NUS-WIDE, Open Images and MS COCO. Furthermore, we show the generalization capabilities of our fusion approach in the zero-shot detection task on MS COCO, achieving favorable performance against existing methods.

Abstract: Multi-domain recommender systems benefit from cross-domain representation learning and positive knowledge transfer. Both can be achieved by introducing a specific modeling of input data (i.e. disjoint history) or trying dedicated training regimes. At the same time, treating domains as separate input sources becomes a limitation as it does not capture the interplay that naturally exists between domains. In this work, we efficiently learn multi-domain representation of sequential users’ interactions using graph neural networks. We use temporal intra- and inter-domain interactions as contextual information for our method called MAGRec (short for Multi-dom Ain Graph-based Recommender). To better capture all relations in a multi-domain setting, we learn two graph-based sequential representations simultaneously: domain-guided for recent user interest, and general for long-term interest. This approach helps to mitigate the negative knowledge transfer problem from multiple domains and improve overall representation. We perform experiments on publicly available datasets in different scenarios where MAGRec consistently outperforms state-of-the-art methods. Furthermore, we provide an ablation study and discuss further extensions of our method.

Abstract: Much of the neural machinery of the early visual cortex, from the extraction of local orientations to contextual modulations through lateral interactions, is thought to have developed to provide a sparse encoding of contour in natural scenes, allowing the brain to process efficiently most of the visual scenes we are exposed to. Certain visual stimuli, however, cause visual stress, a set of adverse effects ranging from simple discomfort to migraine attacks, and epileptic seizures in the extreme, all phenomena linked with an excessive metabolic demand. The theory of efficient coding suggests a link between excessive metabolic demand and images that deviate from natural statistics. Yet, the mechanisms linking energy demand and image spatial content in discomfort remain elusive. Here, we used theories of visual coding that link image spatial structure and brain activation to characterize the response to images observers reported as uncomfortable in a biologically based neurodynamic model of the early visual cortex that included excitatory and inhibitory layers to implement contextual influences. We found three clear markers of aversive images: a larger overall activation in the model, a less sparse response, and a more unbalanced distribution of activity across spatial orientations. When the ratio of excitation over inhibition was increased in the model, a phenomenon hypothesised to underlie interindividual differences in susceptibility to visual discomfort, the three markers of discomfort progressively shifted toward values typical of the response to uncomfortable stimuli. Overall, these findings propose a unifying mechanistic explanation for why there are differences between images and between observers, suggesting how visual input and idiosyncratic hyperexcitability give rise to abnormal brain responses that result in visual stress.