Over the recent years, Reinforcement Learning combined with Deep Learning techniques has successfully proven to solve complex problems in various domains, including robotics, self-driving cars, and finance. In this article, we are introducing Reinforcement Learning (RL) to label placement, a complex task in data visualization that seeks optimal positioning for labels to avoid overlap and ensure legibility. Our novel point-feature label placement method utilizes Multi-Agent Deep Reinforcement Learning to learn the label placement strategy, the first machine-learning-driven labeling method, in contrast to the existing hand-crafted algorithms designed by human experts. To facilitate RL learning, we developed an environment where an agent acts as a proxy for a label, a short textual annotation that augments visualization. Our results show that the strategy trained by our method significantly outperforms the random strategy of an untrained agent and the compared methods designed by human experts in terms of completeness (i.e., the number of placed labels). The trade-off is increased computation time, making the proposed method slower than the compared methods. Nevertheless, our method is ideal for scenarios where the labeling can be computed in advance, and completeness is essential, such as cartographic maps, technical drawings, and medical atlases. Additionally, we conducted a user study to assess the perceived performance. The outcomes revealed that the participants considered the proposed method to be significantly better than the other examined methods. This indicates that the improved completeness is not just reflected in the quantitative metrics but also in the subjective evaluation by the participants.

In this article, we present an algorithm capable of mixed labeling of 2D and 3D objects. In mixed labeling, the given objects are labeled with both internal labels placed (at least partially) over the objects and external labels placed in the space around the objects and connected with the labeled objects with straight-line leaders. The proposed algorithm determines the position and type of each label based on the user-specified ambiguity threshold and eliminates overlaps between the labels, as well as between the internal labels and the straight-line leaders of external labels. The algorithm is a screen-space technique; it operates in an image where the 2D objects or projected 3D objects are encoded. In other words, we can use the algorithm whenever we can render the objects to an image, which makes the algorithm fit for use in many domains. The algorithm operates in real-time, giving the results immediately. Finally, we present results from an expert evaluation, in which a professional illustrator has evaluated the label layouts produced with the proposed algorithm.

Labels, short textual annotations are an important component of data visualizations, illustrations, infographics, and geographical maps. In interactive applications, the labeling method responsible for positioning the labels should not take the resources from the application itself. In other words, the labeling method should provide the result as fast as possible. In this work, we propose a greedy point-feature labeling method running on GPU. In contrast to existing methods that position the labels sequentially, the proposed method positions several labels in parallel. Yet, we guarantee that the positioned labels will not overlap, nor will they overlap important visual features. When the proposed method is searching for the label position of a point-feature, the available label candidates are evaluated with respect to overlaps with important visual features, conflicts with label candidates of other point-features, and their ambiguity. The evaluation of each label candidate is done in constant time independently from the number of point-features, the number of important visual features, and the resolution of the created image. Our measurements indicate that the proposed method is able to position more labels than existing greedy methods that do not evaluate conflicts between the label candidates. At the same time, the proposed method achieves a significant increase in performance. The increase in performance is mainly due to the parallelization and the efficient evaluation of label candidates.

We present a new algorithm for calculating the external labeling of ghosted views of moderately complex 3D models. The algorithm uses multiple criteria decision making, based on fuzzy logic, to optimize positions of the labels associated with different parts of the input model. The proposed method can be used with various existing algorithms for creating ghosted views from 3D models. The method operates in real-time, which allows the user to acquire a good understanding of the structure of the input model by studying the model and its labels from different viewpoints. We have conducted a user study to evaluate label layouts produced by our algorithm and those created by humans. The results show that the proposed method can significantly improve user understanding of labeled ghosted views of complicated 3D models, and its label layouts are comparable with label layouts created by humans.

Labeling is intrinsically important for exploring and understanding complex environments and models in a variety of domains. We present a method for interactive labeling of crowded 3D scenes containing very many instances of objects spanning multiple scales in size. In contrast to previous labeling methods, we target cases where many instances of dozens of types are present and where the hierarchical structure of the objects in the scene presents an opportunity to choose the most suitable level for each placed label. Our solution builds on and goes beyond labeling techniques in medical 3D visualization, cartography, and biological illustrations from books and prints. In contrast to these techniques, the main characteristics of our new technique are: 1) a novel way of labeling objects as part of a bigger structure when appropriate, 2) visual clutter reduction by labeling only representative instances for each type of an object, and a strategy of selecting those. The appropriate level of label is chosen by analyzing the scene's depth buffer and the scene objects' hierarchy tree. We address the topic of communicating the parent-children relationship between labels by employing visual hierarchy concepts adapted from graphic design. Selecting representative instances considers several criteria tailored to the character of the data and is combined with a greedy optimization approach. We demonstrate the usage of our method with models from mesoscale biology where these two characteristics-multi-scale and multi-instance-are abundant, along with the fact that these scenes are extraordinarily dense.

We present an external labeling laid over small and/or overlapping 2D objects as an efficient representation for their selection. The approximation of objects with points allows us to transform the labeling problem to graph layout problem, which we solve by means of force-based algorithm. The input parameters allow us to influence the resulting layout of label boxes (e.g. to adapt their distance for imprecise input devices). In a study with 15 participants two implementations of our algorithm were compared against labeling method, where all label boxes share the same offset from corresponding objects. The results of the study show that implementation using a special functionality (temporary freezing of the label box position recalculation) was 14% faster with a comparable accuracy. The subjective evaluation revealed that the implementation with temporary freezing is perceived as most comfortable, fastest and most accurate. The implementation without temporary freezing showed much higher error rate and cannot be recommended.