Tracking devices that can track both players and balls are critical to the performance of sports teams. Recently, significant effort has been focused on building larger broadcast sports video datasets. However, broadcast videos do not show the entire pitch and only provides partial information about the game. On the other hand, other camera perspectives can capture the whole field in a single frame, such as fish-eye and bird-eye view (drone) cameras. Unfortunately, there has not been a dataset where such data has been publicly shared until now. This paper proposes SoccerTrack, a dataset set consisting of GNSS and bounding box tracking data annotated on video captured with a 8K-resolution fish-eye camera and a 4K-resolution drone camera. In addition to a benchmark tracking algorithm, we include code for camera calibration and other preprocessing. Finally, we evaluate the tracking accuracy among a GNSS, fish-eye camera and drone camera data. SoccerTrack is expected to provide a more robust foundation for designing MOT algorithms that are less reliant on visual cues and more reliant on motion analysis. The dataset and related project code is available at https://github.com/AtomScott/SoccerTrack¹².

Hyperparameter optimization (HPO) is a vital step in improving performance in deep learning (DL). Practitioners are often faced with the trade-off between multiple criteria, such as accuracy and latency. Given the high computational needs of DL and the growing demand for efficient HPO, the acceleration of multi-objective (MO) optimization becomes ever more important. Despite the significant body of work on meta-learning for HPO, existing methods are inapplicable to MO tree-structured Parzen estimator (MO-TPE), a simple yet powerful MO-HPO algorithm. In this paper, we extend TPE’s acquisition function to the meta-learning setting using a task similarity defined by the overlap of top domains between tasks. We also theoretically analyze and address the limitations of our task similarity. In the experiments, we demonstrate that our method speeds up MO-TPE on tabular HPO benchmarks and attains state-of-the-art performance. Our method was also validated externally by winning the AutoML 2022 competition on “Multiobjective Hyperparameter Optimization for Transformers”. See https://arxiv.org/abs/2212.06751 for the latest version with Appendix.

We developed a robust object-level change detection method that could capture distinct scene changes in an image pair with viewpoint differences. To achieve this, we designed a network that could detect object-level changes in an image pair. In contrast to previous studies, we considered the change detection task as a graph matching problem for two object graphs that were extracted from each image. By virtue of this, the proposed network more robustly detected object-level changes with viewpoint differences than existing pixel-level approaches. In addition, the network did not require pixel-level change annotations, which have been required in previous studies. Specifically, the proposed network extracted the objects in each image using an object detection module and then constructed correspondences between the objects using an object matching module. Finally, the network detected objects that appeared or disappeared in a scene using the correspondences that were obtained between the objects. To verify the effectiveness of the proposed network, we created a synthetic dataset of images that contained object-level changes. In experiments on the created dataset, the proposed method improved the ${\mathrm{F}}_1$ score of conventional methods by more than 40%. Our synthetic dataset will be available publicly online.

Mode choice models are important for investigating how travelers will react to changes in public transportation fares and the introduction of new mobility services. The models are essentially built for a given set of mode alternatives, for which parametric utility functions for each mode are defined, and the parameters are estimated using mode choice behavior data. Therefore, the models are dependent on the target mode alternatives in the modeling step and not generalizable to other modes. This study aimed to develop a general mode choice model that can be applied to various sets of mode alternatives. We used large language models to achieve generalizability. The model input comprised sentences that represented different alternatives and variables related to choosing a travel mode. The output was a word that indicated the mode that would be selected. In this study, we created a textual dataset based on four publicly available mode choice datasets. The experimental results showed that the proposed language-based mode choice model, our proposed approach, was more versatile than the classical multinomial logit model in predicting a variety of mode alternative sets.

In this study, we propose an improved version of Pareto deep Q-network (PDQN), a multi-objective deep reinforcement learning method, and attempt to demonstrate its effectiveness in a real-world problem such as crowd route guidance strategy optimization. Overcrowding during crowd movement can sometimes lead to accidents; therefore, it is imperative to guide crowds to move safely and efficiently. Safety and efficiency are conflicting objectives, and how to dynamically determine guidance can be formulated as a multi-objective sequential decision-making problem. PDQN is suitable for solving these problems, but its applicability to complex real-world problems has not been fully verified. To apply PDQN to real problems, we propose to adjust the parameters of PDQN and improve the action selection criteria during learning. A toy problem and a crowd guidance problem using multi-agent crowd simulation are adopted to evaluate the performance of improved PDQN. Experimental results show that the improved PDQN can search for good strategies compared with the original PDQN and can obtain better strategies than simplified strategies.