Explainability and interpretability are two critical aspects of decision support systems. Within computer vision, they are critical in certain tasks related to human behavior analysis such as in health care applications. Despite their importance, it is only recently that researchers are starting to explore these aspects. This paper provides an introduction to explainability and interpretability in the context of computer vision with an emphasis on looking at people tasks. Specifically, we review and study those mechanisms in the context of first impressions analysis. To the best of our knowledge, this is the first effort in this direction. Additionally, we describe a challenge we organized on explainability in first impressions analysis from video. We analyze in detail the newly introduced data set, evaluation protocol, proposed solutions and summarize the results of the challenge. Finally, derived from our study, we outline research opportunities that we foresee will be decisive in the near future for the development of the explainable computer vision field.Keywords Explainable computer vision · First impressions · Personality analysis · Multimodal information · Algorithmic accountability 1 IntroductionLooking at People (LaP) -the field of research focused on the visual analysis of human behavior -has been a very active research field within computer vision in the last decade [28,29,62]. Initially, LaP focused on tasks associated with basic human behaviors that were obviously visual (e.g., basic gesture recognition [71,70] or face recognition in restricted scenarios [10,83]). Research progress in LaP has now led to models that can solve those initial tasks relatively easily [66,82]. Instead, attention on human behavior analysis has now turned to problems that are not visually evident to model / recognize [84,48,72]. For instance, consider the task of assessing personality traits from visual information [72]. Although there are methods that can estimate apparent personality traits with (relatively) acceptable performance, model recommendations by themselves are useless if the end user is not confident on the model's reasoning, as the primary use for such estimation is to understand bias in human assessors.Explainability and interpretability are thus critical features of decision support systems in some LaP tasks [26]. The former focuses on mechanisms that can tell what is the rationale behind the decision or recommendation made by

No abstract

Abstract. In the past few years, a lot of attention has been devoted to multimedia indexing by fusing multimodal informations. Two kinds of fusion schemes are generally considered: The early fusion and the late fusion. We focus on late classifier fusion, where one combines the scores of each modality at the decision level.To tackle this problem, we investigate a recent and elegant well-founded quadratic program named MinCq coming from the machine learning PAC-Bayesian theory. MinCq looks for the weighted combination, over a set of real-valued functions seen as voters, leading to the lowest misclassification rate, while maximizing the voters' diversity. We propose an extension of MinCq tailored to multimedia indexing. Our method is based on an order-preserving pairwise loss adapted to ranking that allows us to improve Mean Averaged Precision measure while taking into account the diversity of the voters that we want to fuse. We provide evidence that this method is naturally adapted to late fusion procedures and confirm the good behavior of our approach on the challenging PASCAL VOC'07 benchmark.

No abstract

International audienceThis paper reviews and discusses research advances on “explainable machine learning” in computer vision. We focus on a particular area of the “Looking at People” (LAP) thematic domain: first impressions and personality analysis. Our aim is to make the computational intelligence and computer vision communities aware of the importance of developing explanatory mechanisms for computer-assisted decision making applications, such as automating recruitment. Judgments based on personality traits are being made routinely by human resource departments to evaluate the candidates' capacity of social insertion and their potential of career growth. However, inferring personality traits and, in general, the process by which we humans form a first impression of people, is highly subjective and may be biased. Previous studies have demonstrated that learning machines can learn to mimic human decisions. In this paper, we go one step further and formulate the problem of explaining the decisions of the models as a means of identifying what visual aspects are important, understanding how they relate to decisions suggested, and possibly gaining insight into undesirable negative biases. We design a new challenge on explainability of learning machines for first impressions analysis. We describe the setting, scenario, evaluation metrics and preliminary outcomes of the competition. To the best of our knowledge this is the first effort in terms of challenges for explainability in computer vision. In addition our challenge design comprises several other quantitative and qualitative elements of novelty, including a “coopetition” setting, which combines competition and collaboration

Concept indexing in multimedia libraries is very useful for users searching and browsing but it is a very challenging research problem as well. Combining several modalities, features or concepts is one of the key issues for bridging the gap between signal and semantics. In this paper, we present three fusion schemes inspired from the classical early and late fusion schemes. First, we present a kernel-based fusion scheme which takes advantage of the kernel basis of classifiers such as SVMs. Second, we integrate a new normalization process into the early fusion scheme. Third, we present a contextual late fusion scheme to merge classification scores of several concepts. We conducted experiments in the framework of the official TRECVID'06 evaluation campaign and we obtained significant improvements with the proposed fusion schemes relatively to usual fusion schemes.

Automatic semantic classification of image databases is very useful for users searching and browsing, but it is at the same time a very challenging research problem as well. Local features based image classification is one of the key issues to bridge the semantic gap in order to detect concepts. This paper proposes a framework for incorporating contextual information into the concept detection process. The proposed method combines local and global classifiers with stacking, using SVM. We studied the impact of topologic and semantic contexts in concept detection performance and proposed solutions to handle the large amount of dimensions involved in classified data. We conducted experiments on TRECVID'04 subset with 48104 images and 5 concepts. We found that the use of context yields a significant improvement both for the topologic and semantic contexts.

In this paper, we compare active learning strategies for indexing concepts in video shots. Active learning is simulated using subsets of a fully annotated data set instead of actually calling for user intervention. Training is done using the collaborative annotation of 39 concepts of the TRECVID 2005 campaign. Performance is measured on the 20 concepts selected for the TRECVID 2006 concept detection task. The simulation allows exploring the effect of several parameters: the strategy, the annotated fraction of the data set, the size of the data set, the number of iterations and the relative difficulty of concepts.Three strategies were compared. The first two, respectively, select the most probable and the most uncertain samples. The third one is a random choice. For easy concepts, the ''most probable'' strategy is the best one when less than 15% of the data set is annotated and the ''most uncertain'' strategy is the best one when 15% or more of the data set is annotated. The ''most probable'' and ''most uncertain'' strategies are roughly equivalent for moderately difficult and difficult concepts. In all cases, the maximum performance is reached when 12-15% of the whole data set is annotated. This result is, however, dependent upon the step size and the training set size. One-fortieth of the training set size is a good value for the step size. The size of the subset of the training set that has to be annotated in order to reach the maximum achievable performance varies with the square root of the training set size. The ''most probable'' strategy is more ''recall oriented'' and the ''most uncertain'' strategy is more ''precision oriented''. r