SoccerNet 2022 Challenges Results

The SoccerNet 2022 challenges were the second annual video understanding challenges organized by the SoccerNet team. In 2022, the challenges were composed of 6 vision-based tasks: (1) action spotting, focusing on retrieving action timestamps in long untrimmed videos, (2) replay grounding, focusing on retrieving the live moment of an action shown in a replay, (3) pitch localization, focusing on detecting line and goal part elements, (4) camera calibration, dedicated to retrieving the intrinsic and extrinsic camera parameters, (5) player re-identification, focusing on retrieving the same players across multiple views, and (6) multiple object tracking, focusing on tracking players and the ball through unedited video streams. Compared to last year's challenges, tasks (1-2) had their evaluation metrics redefined to consider tighter temporal accuracies, and tasks (3-6) were novel, including their underlying data and annotations. More information on the tasks, challenges and leaderboards are available on https://www.soccer-net.org. Baselines and development kits are available on https://github.com/SoccerNet.

Sports Re-ID: Improving Re-Identification Of Players In Broadcast Videos Of Team Sports

This work focuses on player re-identification in broadcast videos of team sports. Specifically, we focus on identifying the same player in images captured from different camera viewpoints during any given moment of a match. This task differs from traditional applications of person re-id in a few important ways. Firstly, players from the same team wear highly similar clothes, thereby making it harder to tell them apart. Secondly, there are only a few number of samples for each identity, which makes it harder to train a re-id system. Thirdly, the resolutions of the images are often quite low and vary a lot. This combined with heavy occlusions and fast movements of players greatly increase the challenges for re-id. In this paper, we propose a simple but effective hierarchical data sampling procedure and a centroid loss function that, when used together, increase the mean average precision (mAP) by 7 - 11.5 and the rank-1 (R1) by 8.8 - 14.9 without any change in the network or hyper-parameters used. Our data sampling procedure improves the similarity of the training and test distributions, and thereby aids in creating better estimates of the centroids of the embeddings (or feature vectors). Surprisingly, our study shows that in the presence of severely limited data, as is the case for our application, a simple centroid loss function based on euclidean distances significantly outperforms the popular triplet-centroid loss function. We show comparable improvements for both convolutional networks and vision transformers. Our approach is among the top ranked methods in the SoccerNet Re-Identification Challenge 2022 leaderboard (test-split) with a mAP of 86.0 and a R1 of 81.5. On the sequestered challenge split, we achieve an mAP of 84.9 and a R1 of 80.1. Research on re-id for sports-related applications is very limited and our work presents one of the first discussions in the literature on this.

Semantic Labeling of Large-Area Geographic Regions Using Multi-View and Multi-Date Satellite Images, and Noisy OSM Training Labels

We present a novel multi-view training framework and CNN architecture for combining information from multiple overlapping satellite images and noisy training labels derived from OpenStreetMap (OSM) for semantic labeling of buildings and roads across large geographic regions (100 km$^2$). Our approach to multi-view semantic segmentation yields a 4-7% improvement in the per-class IoU scores compared to the traditional approaches that use the views independently of one another. A unique (and, perhaps, surprising) property of our system is that modifications that are added to the tail-end of the CNN for learning from the multi-view data can be discarded at the time of inference with a relatively small penalty in the overall performance. This implies that the benefits of training using multiple views are absorbed by all the layers of the network. Additionally, our approach only adds a small overhead in terms of the GPU-memory consumption even when training with as many as 32 views per scene. The system we present is end-to-end automated, which facilitates comparing the classifiers trained directly on true orthophotos vis-a-vis first training them on the off-nadir images and subsequently translating the predicted labels to geographical coordinates. With no human supervision, our IoU scores for the buildings and roads classes are 0.8 and 0.64 respectively which are better than state-of-the-art approaches that use OSM labels and that are not completely automated.

A Comparative Evaluation of SGM Variants (including a New Variant, tMGM) for Dense Stereo Matching

Our goal here is threefold: [1] To present a new dense-stereo matching algorithm, tMGM, that by combining the hierarchical logic of tSGM with the support structure of MGM achieves 6-8\% performance improvement over the baseline SGM (these performance numbers are posted under tMGM-16 in the Middlebury Benchmark V3 ); and [2] Through an exhaustive quantitative and qualitative comparative study, to compare how the major variants of the SGM approach to dense stereo matching, including the new tMGM, perform in the presence of: (a) illumination variations and shadows, (b) untextured or weakly textured regions, (c) repetitive patterns in the scene in the presence of large stereo rectification errors. [3] To present a novel DEM-Sculpting approach for estimating initial disparity search bounds for multi-date satellite stereo pairs. Based on our study, we have found that tMGM generally performs best with respect to all these data conditions. Both tSGM and MGM improve the density of stereo disparity maps and combining the two in tMGM makes it possible to accurately estimate the disparities at a significant number of pixels that would otherwise be declared invalid by SGM. The datasets we have used in our comparative evaluation include the Middlebury2014, KITTI2015, and ETH3D datasets and the satellite images over the San Fernando area from the MVS Challenge dataset.

A New Stereo Benchmarking Dataset for Satellite Images

In order to facilitate further research in stereo reconstruction with multi-date satellite images, the goal of this paper is to provide a set of stereo-rectified images and the associated groundtruthed disparities for 10 AOIs (Area of Interest) drawn from two sources: 8 AOIs from IARPA's MVS Challenge dataset and 2 AOIs from the CORE3D-Public dataset. The disparities were groundtruthed by first constructing a fused DSM from the stereo pairs and by aligning 30 cm LiDAR with the fused DSM. Unlike the existing benckmarking datasets, we have also carried out a quantitative evaluation of our groundtruthed disparities using human annotated points in two of the AOIs. Additionally, the rectification accuracy in our dataset is comparable to the same in the existing state-of-the-art stereo datasets. In general, we have used the WorldView-3 (WV3) images for the dataset, the exception being the UCSD area for which we have used both WV3 and WorldView-2 (WV2) images. All of the dataset images are now in the public domain. Since multi-date satellite images frequently include images acquired in different seasons (which creates challenges in finding corresponding pairs of pixels for stereo), our dataset also includes for each image a building mask over which the disparities estimated by stereo should prove reliable. Additional metadata included in the dataset includes information about each image's acquisition date and time, the azimuth and elevation angles of the camera, and the intersection angles for the two views in a stereo pair. Also included in the dataset are both quantitative and qualitative analyses of the accuracy of the groundtruthed disparity maps. Our dataset is available for download at \url{https://engineering.purdue.edu/RVL/Database/SatStereo/index.html}