Live indoor sports broadcasts require dynamic camera control in response to relevant game situations such as penalties or timeouts, a process that traditionally relies on human operators. This paper presents a solution for automatic real-time classification of game states in indoor invasion sports, with handball as the primary case study and basketball as a secondary validation scenario. Our approach utilizes raw video directly from cameras, enabling real-time analysis. The system continuously processes video frames, assigning each to one of seven classes: left/right attack, left/right counterattack, left/right penalty, and timeout. A synthetic representation of each frame is used to standardize the depiction of game dynamics. The proposed pipeline includes object detection with a fine-tuned You Only Look Once (YOLO) model to locate players, the ball, and referees; object tracking to compute velocity vectors; generation of a synthetic frame representing the current game state; and final classification using a custom Dense Convolutional Network (DenseNet). Using a dataset of 20 handball matches, the proposed system achieved a macro-averaged F1-score of 96.1%, with a per-image inference time below 4 milliseconds, evaluated on 118,129 images from matches unseen during training. The same pipeline was subsequently applied to basketball using only two matches, achieving an F1-score of 92.5% on 12,390 images, thereby illustrating the transferability of the proposed approach to other indoor invasion sports. The full pipeline operates in 34.04 milliseconds with GPU acceleration, processing over 25 frames per second.

In the era of social media and instantaneous content consumption, the demand for quick post-event sports highlights has emerged, requiring systems to deliver compelling summaries in a minimal turnaround time. This paper introduces an AI-driven framework designed to streamline the creation of match recaps by leveraging real-time data acquisition and efficient post-processing. The system analyzes multi-modal inputs —including live game statistics, audio-visual feeds, and contextual cues— to automatically identify key moments (e.g., goals, pivotal plays) as they occur. By integrating lightweight neural models and rule-based prioritization, it generates timestamped clips immediately after the match concludes, significantly reducing manual editing effort. The solution not only supports human editors by providing pre-curated material but also enables fully automated highlight production for platforms requiring instant content delivery. Evaluations on soccer and basketball matches demonstrate the system’s ability to cut post-event processing time by 87.5% while maintaining 90% accuracy in event selection compared to manual curation. The work underscores the potential of hybrid AI systems to bridge real-time analytics with post-production workflows, offering scalability across sports and media formats. The generated highlights are already being published on a production platform (https://tiivii.gal), demonstrating real-world applicability.

Abstract Real-time play recognition and classification algorithms are crucial for automating video production and live broadcasts of sporting events. However, current methods relying on human pose estimation and deep neural networks introduce high latency on commodity hardware, limiting their usability in low-cost real-time applications. We present PlayNet, a novel approach to real-time handball play classification. Our method is based on Kalman embeddings, a new low-dimensional representation for game states that enables efficient operation on commodity hardware and customized camera layouts. Firstly, we leverage Kalman filtering to detect and track the main agents in the playing field, allowing us to represent them in a single normalized coordinate space. Secondly, we utilize a neural network trained in nonlinear dimensionality reduction through fuzzy topological data structure analysis. As a result, PlayNet achieves real-time play classification with under 55 ms of latency on commodity hardware, making it a promising addition to automated live broadcasting and game analysis pipelines.

During the broadcast of sporting events, certain situations such as a penalty or a time-out occur, for which a specific action is required. In traditional broadcasting, many people are implied in making decisions based on what is happening at any given moment. To broadcast quality and entirely automatically matches it is necessary to be able to classify the important situations and then make decisions based on them. This paper presents a solution based on deep learning which is able to classify the main states of a handball match. The generated model has been trained using 127,600 images of 13 local team matches. On a test set of 118,129 images of other 7 matches, it is able to classify these situations with an accuracy of 98.6% in only 4 milliseconds, allowing to analyze the state of the game in real time. The full pipeline takes only 34.04 milliseconds using GPU acceleration, processing more than 25 frames per seconds.

Image mapping on 3D huge scenarios of the real world is one of the most fundamental and computational expensive processes for the integration of multi-source sensing data. Recent studies focused on the observation and characterization of Earth have been enhanced by the proliferation of Unmanned Aerial Vehicle (UAV) and sensors able to capture massive datasets with a high spatial resolution. Despite the advances in manufacturing new cameras and versatile platforms, only a few methods have been developed to characterize the study area by fusing heterogeneous data such as thermal, multispectral or hyperspectral images with high-resolution 3D models. The main reason for this lack of solutions is the challenge to integrate multi-scale datasets and high computational efforts required for image mapping on dense and complex geometric models. In this paper, we propose an efficient pipeline for multi-source image mapping on huge 3D scenarios. Our GPU-based solution significantly reduces the run time and allows us to generate enriched 3D models on-site. The proposed method is out-of-core and it uses available resources of the GPU’s machine to perform two main tasks: (i) image mapping and (ii) occlusion testing. We deploy highly-optimized GPU-kernels for image mapping and detection of self-hidden geometry in the 3D model, as well as a GPU-based parallelization to manage the 3D model considering several spatial partitions according to the GPU capabilities. Our method has been tested on 3D scenarios with different point cloud densities (66M, 271M, 542M) and two sets of multispectral images collected by two drone flights. We focus on launching the proposed method on three platforms: (i) System on a Chip (SoC), (ii) a user-grade laptop and (iii) a PC. The results demonstrate the method’s capabilities in terms of performance and versatility to be computed by commodity hardware. Thus, taking advantage of GPUs, this method opens the door for embedded and edge computing devices for 3D image mapping on large-scale scenarios in near real-time.

The hierarchical radiosity algorithm provides high-quality illumination and the view-independent results obtained can be re-employed for different camera positions. On the other hand, the utilization of multiresolution models is a common solution for the real-time rendering of complex scenes. In this case, the level of detail of each object depends on the specific camera position. Unfortunately, global illumination computations may present an important performance loss in this context. In this paper we present a new solution for hierarchical radiosity for multiresolution systems. Our proposal is based on the application of an enriched hierarchical radiosity algorithm to an input scene with low-resolution objects (represented by coarse meshes) and the efficient data management of the resulting values. The representation of the information we use permits the application of the radiosity values obtained for the coarse version of an object to finer resolution versions of that object. Results of our implementation show that our algorithm produces high-quality images with an important reduction in computational costs.

Physically-based illumination is an essential factor for realistic rendering. In this context, hierarchical radiosity is one of the most accurate global illumination methods. One of the key features of the radiosity approach is that it obtains view-independent global illumination results. Unfortunately, global illumination has high computational and memory requirements, and hierarchical radiosity, though more efficient than other radiosity solutions, is not an exception. The progressive popularization of multiprocessor and multi-core processor systems makes the design and implementation of efficient parallel algorithms an appealing alternative in this field. In this paper we present a novel parallel radiosity method addressing the hierarchical radiosity computation on current homogeneous multi-core environments. One of the main contributions of our work is the use of different tasks to exploit the independent interactions among the geometric elements in the scene. Our parallel solution leads to a versatile radiosity implementation that takes advantage of the multiple computational resources in the system, such as multi-core processors and SMT (Simultaneous Multithreading) capabilities. Good results in terms of performance have been achieved.

Nowadays, one of the most active areas of research in computer graphics is focused on improving the performance of global illumination methods. In this paper we present a new parallel method for hierarchical radiosity computation. Contrary to usual solutions in the literature our approach carries out an effective partitioning and distribution of the input geometry among the processors. Thus, the memory constraint is reduced when processing large scenes. The method presented in this paper also introduces a multi-thread scheduling to improve interaction between processors. This solution replaces the polling strategy usually used in parallel computing, when several independent tasks need to establish asynchronous communication. The great flexibility of our parallel implementation has been demonstrated by the good results achieved on a testbed with different computing systems and test scenes

The demand for high quality and speed in graphical representations of 3D scenes is continuously increasing. An important aspect of visual realism is real-time processing of transparent objects. In order to provide a realistic solution to be considered for future graphics cards, proposed extensions to the hardware need to be small. In this paper, we propose two new hardware oriented proposals for computing order independent transparency for any number of layers. The first proposal is based on the precomputation of the contribution factor of each element to the final colour of the pixel. This permits the evaluation of the transparent objects without requiring their previous sorting. The second proposal is based on the utilization of a pre-sorting stage so that the transparent objects are organized in a back to front order. In both proposals, time requirements are reduced with respect to previous solutions.