Optimizing Through Learned Errors for Accurate Sports Field Registration

Extraction of Positional Player Data from Broadcast Soccer Videos

“…We extend the pan and tilt range to (−40 • , 40 • ) and (−20 • , −5 • ), respectively, in all models. As the tactic-cam obviously covers a wider range (especially focal length as seen in Figure 2 A,D,E), we also WC14 [16] TV12 TV14 TC14 Results: As reported in Table 2 the reproduced results (base parameters) from Chen and Little [4] at WC14 are of similar quality compared to other methods [19,37,46]. We observe a noticeable drop in IoU part on our test sets where the camera parameters (especially the camera position (x, y, z)) are unknown.…”

“…Sharma et al [47] and Chen and Little [4] propose the nearest neighbor search from a synthetic dataset of pairs of edge images and camera images for fully-automated registration. Jiang et al [19] present a two-step deep learning approach that initially estimates a homography and minimizes the error using another deep network instead of the Lucas-Kanade algorithm [1]. Citraro et al [6] suggest an approach that also takes into account the position of players and is trained on a separate dataset for uncalibrated cameras.…”

“…The camera calibration is defined as the nearest neighbor search in a synthetic dataset of edge map camera pairs. We choose this approach for multiple reasons: (1) It obtains almost state-of-the-art performance on the only test set for soccer [16], (2) it does not rely on manual annotations to obtain training data [5,19,37], and (3) is adaptable to other environments (e.g., stadiums and camera parameters) by changing only a few hyper-parameters, as shown in our experiments (Section 4.4).…”

“…While the test set follows the same data distribution as the training data, in particular, the camera hyper-parameters (location, focal-length, etc. ), capabilities are not investigated by existing solutions [4,19,37,46]. Primarily, to indicate the adaptability of Chen and Little [4]'s approach (Section 3.2) to different environmental settings, we explore several hyper-parameters on our target test sets (see Section 4.1).…”

Evaluating Mixed-Precision Arithmetic for 3D Generative Adversarial Networks to Simulate High Energy Physics Detectors

“…Following the progress made in image classification tasks [1], [2] by Khrizhevsky et al [3], Szegedy et al [4] and He et al [5], Deep Neural Networks (DNNs) [6] have been successfully applied to domains like biology [7], economy [8], chemistry [9], or sports management [10]. Many refined techniques have emerged to tackle these new problems beyond image classification.…”