Synthesizing realistic football team formations is a high-dimensional coordination problem that balances rigid physical constraints with fluid tactical objectives. We frame this problem as a 22-dimensional continuous black-box optimization task, where the goal is to optimize 11-player (x, y) coordinates tailored to specific phases of play: attacking possession, defensive possession, and defensive. We introduce the first application of Covariance Matrix Adaptation Evolution Strategies (CMA-ES) to phase-conditioned tactical pattern synthesis, leveraging its covariance adaptation to capture the intricate spatial dependencies between players. Benchmarked against a Differential Evolution (DE) baseline and validated using empirical tracking data, our results show that CMA-ES (in a static scenario where the opponent configuration is fixed) consistently converges to professional-grade tactical structures. Furthermore, while introducing reactive opponent behavior creates a challenging non-stationary landscape, the CMA-ES approach significantly outperforms the DE baseline in generating coherent, coordinated team formations. This work establishes a new intersection between evolutionary search and automated sports analytics, providing a robust framework for objective tactical simulation.
Underwater image enhancement is challenging due to wavelength dependent attenuation, scattering, color distortion, and loss of local structural details. Recent CNN- and Transformer-based methods have improved restoration quality, but many of them rely on large restoration backbones with high parameter counts and computational cost. This limits their practicality for underwater surveillance, marine robotics, and embedded vision platforms. In this paper, we propose DIVE-Net, a Depthwise Multi-Dilation Vision Network for efficient underwater image enhancement. DIVE-Net follows a compact NAFNet-inspired restoration design, but replaces single-scale depthwise spatial mixing with a lightweight multi-scale depthwise convolution module. The proposed module uses three parallel depthwise branches with dilation rates of 1, 2, and 4, and combines them using learnable softmax weights to aggregate local, intermediate, and wider contextual information. The network is organized as a compact multi-scale encoder–decoder with residual output prediction at multiple resolutions. A learnable output residual scaling factor is further used to control the correction added to each scale output. Experiments on the paired UIEB test set show that DIVE-Net achieves 22.32 dB PSNR and 0.893 SSIM with only 0.74M parameters and 5.49G FLOPs. Compared with substantially larger underwater enhancement models, the proposed method provides a strong accuracy–efficiency trade-off. Results on unpaired U45 and C60 datasets further indicate competitive generalization without dataset-specific fine-tuning.
Unmanned Aerial Vehicles (UAVs) are increasingly deployed in Search and Rescue (SAR) operations because they can rapidly survey remote or difficult terrain and improve situational awareness in time-critical environments. However, the volume and heterogeneity of UAV data can also overload operators and hinder rapid decisions. Recent advances in Multimodal Large Language Models (MLLMs) enable joint reasoning over UAV imagery and mission text, but effective SAR support also requires structured domain knowledge and transparent, auditable outputs that operators can verify and trust. In this work, we propose xC2SAR, an explainable and auditable Command and Control (C2) framework for UAV-supported SAR operations that structures multimodal MLLM reasoning around the formal land-search strategies of the Australian National Search and Rescue Manual (NATSAR). The framework combines structured knowledge retrieval of SAR guidance, UAV imagery interpretation, and concept- and attribution-based explanations within a unified, auditable reasoning chain for human–UAV SAR teaming. We evaluate xC2SAR through a three-layer assessment combining scenario-based operational reasoning, quantitative explanation-quality metrics, and the Assessment List for Trustworthy AI (ALTAI). Code and data will be made available at https://github.com/nikomili/xC2SAR .
