PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Video Action Recognition in Sports: Datasets, Methods and Applications

Fei Wu, Qingzhong Wang, Jian Bian, Haoyi Xiong, Ning Ding, Feixiang Lu, Jun Cheng, Dejing Dou

arXiv:2206.01038v1cs.CVcs.AIcs.MM

TL;DR

Sports video action recognition spans team and individual activities and supports coaching, training, and broadcast-highlight applications. This survey reviews sports datasets and recognition methods, compares their status across sports, and provides a toolbox. It reports that 3D and two-stream models are common, while some datasets remain challenging.

  • Problem

    Sports video action recognition spans individual and team activities, and recognizing player actions matters for coaching, training, and broadcast-highlight localization.

    Source: §I · p. 2, §VII · p. 17

  • Method

    The paper surveys sports datasets and action-recognition methods, compares their performance through literature reviews and experiments, and provides a deep-learning toolbox.

    Source: §I · p. 2, §VII · p. 17

  • Results

    3D and two-stream models are relatively popular; two-stream models achieve 91.4% on TTStroke-21, while NCAA and FineBasketball remain challenging.

    Source: 5) Others · p. 13

  • Takeaways & Limitations

    The toolbox supports video action recognition in multiple sports, including figure skating, football, basketball, and table tennis.

    Source: §I · p. 2

Abstract

from arXiv · show

To understand human behaviors, action recognition based on videos is a common approach. Compared with image-based action recognition, videos provide much more information. Reducing the ambiguity of actions and in the last decade, many works focused on datasets, novel models and learning approaches have improved video action recognition to a higher level. However, there are challenges and unsolved problems, in particular in sports analytics where data collection and labeling are more sophisticated, requiring sport professionals to annotate data. In addition, the actions could be extremely fast and it becomes difficult to recognize them. Moreover, in team sports like football and basketball, one action could involve multiple players, and to correctly recognize them, we need to analyse all players, which is relatively complicated. In this paper, we present a survey on video action recognition for sports analytics. We introduce more than ten types of sports, including team sports, such as football, basketball, volleyball, hockey and individual sports, such as figure skating, gymnastics, table tennis, tennis, diving and badminton. Then we compare numerous existing frameworks for sports analysis to present status quo of video action recognition in both team sports and individual sports. Finally, we discuss the challenges and unsolved problems in this area and to facilitate sports analytics, we develop a toolbox using PaddlePaddle, which supports football, basketball, table tennis and figure skating action recognition.

I. INTRODUCTION

The survey reviews sports video action recognition across team and individual sports, where dataset construction and analysis requirements differ from coarse everyday-action recognition. It maps datasets, methods, applications, and challenges, and introduces a toolbox for sports action recognition.

I. INTRODUCTION

  • Fine-grained sports labels such as figure-skating moves require domain knowledge and professional annotators, unlike coarse everyday-action labels.The introduction identifies specialized annotation as a challenge for sports datasets.Source: §I · p. 1, Figure 1 · p. 1
  • Sports action recognition has applications in coaching, player training, and sports-broadcast highlight generation.Group-activity recognition can assist coaches and players, while individual-action recognition can support correction of small action errors.Source: §I · p. 2
  • Team-sport analysis must account for multiple players, ball trajectories, and player–ball interactions, while individual-sport analysis can focus on one or two players.The survey distinguishes team and individual sports as requiring different analytics.Source: §I · p. 2
  • The survey covers more than ten sports, including football, basketball, volleyball, hockey, diving, tennis, gymnastics, and table tennis.The paper groups its coverage into team and individual sports.Source: §I · p. 2
  • The paper organizes sports genres, maps action-recognition methods, and summarizes sports-related datasets.Table II lists datasets and task types, including detection, classification, tracking, assessment, segmentation, and localization.Source: §I · p. 2, Table II · p. 3
  • The survey reviews current methods and challenges across sports and provides a publicly available deep-learning toolbox for sports action recognition.The related-work discussion distinguishes its sports action-recognition focus from prior surveys and says the toolbox supports figure skating, football, basketball, and table tennis.Source: §I · p. 2

II. SPORTS-RELATED DATASETS

The section outlines how sports-action datasets are assembled and surveys football datasets with different scales, annotation schemes, and uses. It highlights the contrast between large-scale action-recognition resources and datasets suited to player detection.

II. SPORTS-RELATED DATASETS

  • Sports-action datasets are built by defining a sport and its action categories, collecting and processing videos, then annotating labels and, where relevant, action times or object boxes.Annotations may cover trimmed clips with labels or untrimmed videos with action start and end times; AVA also annotates bounding boxes of objects involved in actions.Source: §II · pp. 3–4

A. Football

  • SoccerDB combines 171,191 video segments, player bounding boxes, 11 activity labels, and 17,115 highlights, supporting detection, recognition, localization, and highlight detection.Its annotations include player positions and temporal boundaries for activity segments.Source: A. Football · p. 4
  • SoccerNet contains 500 match videos with three action categories and averages 8.7 actions per hour; SoccerNet-v2 re-labels those videos with 17 categories and denser actions.The section contrasts SoccerNet’s sparse events with SoccerNet-v2’s broader, denser annotations.Source: A. Football · p. 4
  • Large-scale datasets and deep models dominate recent soccer action recognition, while SoccerDB, SSET, and ComprehensiveSoccer are more feasible for tasks requiring player detection.The section distinguishes the resources’ suitability by task rather than treating the datasets as interchangeable.Source: A. Football · p. 4

B. Basketball

The basketball and volleyball datasets vary in scale, annotation, and supported tasks, from player detection to fine-grained action recognition. Several provide player locations alongside action labels, while volleyball ball detection is challenged by fast, blurred motion.

B. Basketball

  • APIDIS combines seven synchronized camera views with player and ball boxes and time-bounded action labels; Basket-1,2 provides synchronized multiview sequences for four ball-related classes.APIDIS supports player detection and action recognition; Basket-1,2 supports action recognition and ball detection.Source: B. Basketball · p. 4
  • NCAA provides 14,548 segments across 14 action categories and 9,000 player-box frames, supporting action recognition and player detection.Its source comprises 257 untrimmed NCAA game videos, typically about 1.5 hours long.Source: B. Basketball · p. 4
  • SpaceJam pairs basketball action labels with estimated player poses, whereas SPIROUDOME annotates player boxes from eight camera views.SpaceJam is suited to skeleton-based action recognition; SPIROUDOME is generally used for player detection.Source: B. Basketball · pp. 4–5
  • FineBasketball offers 26 fine-grained categories across 3,399 segments, while NPUBasketball supplies RGB, depth, and skeleton data for 12 action categories.FineBasketball is imbalanced: crossover dribbling has 717 segments, compared with 12 for follow-up shot.Source: B. Basketball · p. 5

C. Volleyball

  • Volleyball-1,2 has two frame sequences with manually annotated ball boxes, but fast, blurred motion makes ball detection challenging.The sequences contain 10,000 and 19,500 frames, respectively.Source: C. Volleyball · p. 5
  • HierVolleyball and its v2 extension label individual-player and team actions; v2 expands coverage to 4,830 frames from 55 videos and adds player boxes.The original covers 1,525 frames from 15 YouTube videos; v2 includes winpoint as a team-activity category.Source: C. Volleyball · p. 5

D. Hockey

The hockey, tennis, and table-tennis datasets span event recognition, player tracking, and stroke analysis, with varied modalities and annotation granularity. Some support localization or retrieval in addition to action recognition.

D. Hockey

  • Hockey Fight provides 1,000 NHL clips labeled fight or non-fight, while Player Tracklet annotates player and referee boxes with identity labels across 84 broadcast clips.Player Tracklet is intended for player tracking and identification.Source: D. Hockey · p. 5

E. Tennis

  • ACASVA labels hit, non-hit, and serve in six broadcast tennis videos, but provides extracted clip features rather than the original videos.Player positions and action time boundaries are also labeled.Source: E. Tennis · p. 5
  • THETIS contains 1,980 videos across 12 tennis actions and includes depth and skeleton modalities alongside RGB.Its additional data include 1,980 depth videos and 1,217 each of 2D and 3D skeleton videos.Source: E. Tennis · p. 5
  • TenniSet labels six event types in five Olympic-match videos with time boundaries and textual action descriptions, enabling recognition, localization, and retrieval.Descriptions include phrases such as “quick serve is an ace.”Source: E. Tennis · p. 5

F. Table Tennis

  • TTStroke-21 labels 1,378 strokes across 21 categories in 120-FPS egocentric videos; the survey describes it as not challenging, possibly because of the high frame rate.The videos cover 94 hours of games.Source: F. Table Tennis · p. 5
  • SPIN uses 150-FPS video with ball boxes and 30 player-joint heatmaps per frame, supporting ball tracking, pose estimation, and spin prediction.Its collection totals 53 hours and 7.5 million high-resolution frames.Source: F. Table Tennis · p. 5
  • OpenTTGames labels 4,271 ball-bounce, net-hit, or empty events, with frame-level ball coordinates and segmentation masks around each event.The dataset supports semantic segmentation, ball tracking, and event classification.Source: F. Table Tennis · p. 6
  • Stroke Recognition contains 22,111 trimmed videos across 11 stroke categories and is less challenging, with a 21-tree random forest achieving 96.20% accuracy; P2A has 2,721 untrimmed broadcast videos totaling 272 hours.P2A annotations include stroke categories and starting/end-frame indices; professional players, including Olympic table tennis players, confirmed the stroke labels.Source: F. Table Tennis · p. 6

G. Gymnastics

These datasets support fine-grained recognition and localization across gymnastics, badminton, figure skating, and diving, often with temporal or hierarchical annotations. Some also include scores or additional labels for action-quality assessment.

G. Gymnastics

  • FineGym covers 303 videos and about 708 hours, with hierarchical event, set, and element labels plus action and subaction boundaries.Element-level recognition and localization are described as much more challenging than event- or set-level tasks.Source: G. Gymnastics · p. 6

H. Badminton

  • Badminton Olympic contains 10 hour-long singles-match videos with player boxes, point-winning locations, and labeled stroke boundaries across 12 stroke categories.Its annotations support player detection, point localization, action recognition, and localization.Source: H. Badminton · p. 6
  • Stroke Forecasting provides 43,191 trimmed badminton clips in 10 stroke categories and also supports predicting the next stroke from earlier rally strokes.The dataset is described as a recent resource for both action recognition and forecasting.Source: H. Badminton · p. 6

I. Figure skating

  • FSD-10 has 10 figure-skating action categories with 91–233 clips per action and also provides action scores for quality assessment.Its clips range from 3 to 30 seconds.Source: §I · p. 6
  • FineSkating uses hierarchical event and action labels across 46 competition videos, with action boundaries and scores for recognition and quality assessment.The videos are from 2018 and 2019; seven event labels include jump and spin.Source: §I · p. 6
  • MCFS annotates 11,656 segments from 38 competitions with three-level labels and action boundaries, supporting recognition and localization.The collection totals 17.3 hours and 1.7 million frames.Source: §I · p. 6

J. Diving

  • Diving48 contains 18,404 segments across 48 fine-grained categories and is described as having relatively low bias for model evaluation.Each action class combines multiple elements, such as backward take-off and half twist.Source: J. Diving · p. 6
  • MTL-AQA provides 1,412 diving samples labeled with action quality scores, action classes, and textual commentary for quality assessment and recognition.Source: J. Diving · p. 6

K. Multiple Types of Sports

The multiple-sports datasets range from coarse sport classification to fine-grained or multi-person action analysis, while other datasets target specific sports and modalities. Their annotations and recording conditions shape the tasks they support.

K. Multiple Types of Sports

  • Sports-classification datasets label each video by sport and are generally used for coarse classification; Win-Fail instead labels action outcomes in 817 video pairs from multiple domains.Win-Fail uses object movements to recognize whether an action succeeds or fails.Source: K. Multiple Types of Sports · pp. 6–7
  • Olympic contains 800 YouTube videos across 16 categories, while SVW covers 4,100 smartphone videos and 44 actions across 30 sports.Olympic includes occlusions and camera movement; SVW video quality is typically lower than broadcast footage.Source: K. Multiple Types of Sports · p. 7
  • Sports 1M contains about one million YouTube videos across 487 hierarchically labeled categories, supporting coarse-to-fine sports classification.The survey notes that this million-scale dataset can alleviate data hunger in deep learning to some extent.Source: K. Multiple Types of Sports · p. 7
  • MultiSports targets multi-person actions across four team sports, with 66 categories, 3,200 videos, temporal labels, and player bounding boxes.Its annotations support action recognition and temporal and spatial localization.Source: K. Multiple Types of Sports · p. 7

L. Others

  • CVBASE Handball and Squash annotate player trajectories and team or stroke actions, while GolfDB adds player boxes and identity attributes to 1,400 golf-swing segments.FenceNet supplies RGB, 3D skeleton, and depth data for six fencing-action categories performed by expert-level fencers.Source: L. Others · p. 7

III. INDIVIDUAL ACTION RECOGNITION

Individual action recognition covers actions involving one person and spans traditional feature-based pipelines and deep models. Within deep approaches, 2D models combine frame-level features over time, with approaches evolving from fusion and recurrent networks toward temporal modules and transformers.

A. Traditional Models

  • Traditional pipelines extract frame features and classify them; on UCF Sports, GIST reaches 60.0% versus 58.6% for HOG, whose features are biased to backgrounds such as grass fields.The section also describes HOG3D with an MLP or KFDA for tennis, with the latter achieving 84.5% AUC on ACASVA.Source: A. Traditional Models · p. 7
  • Action bank uses appearance-invariant template detectors across scales and views; improved trajectories reach 91.1% on Olympic, versus 62.4% for original trajectories and 82.4% for MBH.MoSIFT combines spatial gradients and optical flow, reaching 89.5% accuracy on Hockey Fight versus 59.0% for STIP.Source: A. Traditional Models · p. 8
  • Hand-crafted spatial-temporal features can perform well, but their calculation is time-consuming and traditional pipelines do not train feature extraction and classification end to end.Deep models are described as more feasible and end-to-end trainable.Source: A. Traditional Models · p. 8, B. Deep Models · p. 8

1) 2D Models:

  • Among four Sports 1M frame-fusion approaches, slow fusion reaches 60.9% accuracy, ahead of single-frame fusion at 59.3%, early fusion at 57.7%, and late fusion at 59.3%.The compared approaches concatenate frame representations, fuse across a time window, or gradually fuse information in higher layers.Source: 1) 2D Models · pp. 8, 10
  • 2D models also combine frame features with LSTMs, sample and aggregate snippets with TSN, or model temporal relations with TRN; TRN reaches 68.7% on FineGym.TSN's RGB-only variant reports 61.4% on FineGym and 87.3% on UCF101; key-frame KTSN reaches 63.3% versus 59.3% on FSD-10.Source: 1) 2D Models · p. 10
  • TSM inserts temporal-shift modules into 2D convolutional blocks, reaching 70.6% on FineGym while using lower computational complexity than the compared models.The passage reports that TSM outperforms 2D TSN, 2D TRN, and some 3D approaches, including I3D.Source: 1) 2D Models · p. 10
  • Transformers are increasingly used for both frame features and temporal modeling, but their larger parameter counts, computational needs, and training difficulty remain constraints.Divided space-time attention reaches 81.0% on Diving48; the section also describes ImageNet pretraining for spatial networks.Source: 1) 2D Models · p. 10

2) 3D Models:

3D models process frame sequences jointly to capture spatial and temporal information. The survey describes a usual performance advantage over 2D models alongside higher computational costs, and reviews designs seeking efficiency or better scaling.

2) 3D Models:

  • Expanding 2D kernels into 3D increases parameter count, computational complexity, and overfitting risk; P3D and (2+1)D designs separate spatial and temporal convolution to address these costs.CSN uses group and depth convolutions and is reported to achieve better performance than 3D CNNs with one third of their FLOPs.Source: 2) 3D Models · pp. 10–11
  • SlowFast uses low-rate and high-rate branches for spatial semantics and object motion, while TPN uses one backbone with a temporal pyramid and matches SlowFast on Kinetics-400.TFCNet adds a fully connected operation over all frame features to SlowFast and reaches 88.3% on Diving48, nearly 11% higher than SlowFast.Source: 2) 3D Models · p. 11
  • 3D transformers and multi-scale designs trade accuracy, data scale, and compute: ViViT's spatial-temporal transformer performs best on large datasets but overfits small ones, while MViT uses around one fifth of ViViT's FLOPs with similar Kinetics-400 performance.ViSwin-large reports 84.9% on Kinetics-400 with ImageNet-21K pretraining; ORViT reaches 88% on Diving48, 8% above its baseline, but requires multi-object tracking.Source: 2) 3D Models · pp. 11–12
  • 3D models are normally superior to 2D models, but can be time-consuming and require more computational resources.The section also notes that pretrain-fine-tune approaches are increasingly popular, particularly for 3D transformers.Source: 2) 3D Models · p. 12

3) Two-stream Models:

Two-stream models combine RGB-frame and optical-flow information through separate neural-network streams. They can outperform one-stream versions, but add optical-flow computation and introduce fusion and overfitting challenges.

3) Two-stream Models:

  • Two-stream models pair RGB frames, which provide spatial and temporal information, with optical flow, which mainly provides motion information.The approach can be applied as a variant of 2D or 3D models.Source: 3) Two-stream Models · p. 12
  • Stream predictions can be averaged, or representations fused using max-pooling, concatenation, bilinear operations, sums, or convolutions.These fusion approaches are described as alternatives for combining frame and optical-flow representations.Source: 3) Two-stream Models · p. 12
  • Two-stream models can outperform one-stream counterparts but may overfit; Gradient Blending estimates loss weights during training to assign a weight to each stream.The passages also note that the streams have different generalizabilities, making a shared training strategy sub-optimal.Source: 3) Two-stream Models · p. 12

4) Skeleton-based Models:

Skeleton-based models use player-joint graphs, commonly modeled with graph convolution, instead of RGB frames. The survey describes speed and appearance trade-offs, noisy joint predictions, and alternatives using joint heatmaps.

4) Skeleton-based Models:

  • ST-GCN models skeleton graphs with graph convolutions, which run faster than frame-based models but perform worse because they cannot capture appearance information.Attention-based variants and EfficientGCN add attention or other techniques; EfficientGCN is reported as competitive on FSD-10 with faster training and greater explainability.Source: 4) Skeleton-based Models · p. 12
  • CTR-GCN models channel-wise topology and reaches 66.2% accuracy on FSD-10, outperforming ST-GCN and AGCN.Its channel-wise topology modeling computes correlations and relationships among graph nodes across channels.Source: 4) Skeleton-based Models · p. 12
  • Joint detection adds computation and prediction noise; PoseC3D uses joint and limb heatmaps to mitigate noise, but remains inferior to two-stream models.The section also reports that skeleton-based approaches overall remain less accurate than other model types.Source: 4) Skeleton-based Models · p. 12
  • The individual-sports comparison covers sports-related datasets outside Table IV; two-stream models reach 91.4% on TTStroke-21, while NCAA and FineBasketball remain challenging.The survey notes that advanced models are rarely used in sports, and that dataset difficulty may help explain this pattern.Source: Table V · p. 13, 5) Others · p. 13

IV. GROUP/TEAM ACTIVITY RECOGNITION

Group activity recognition (GAR) predicts a collective activity from the actions and interactions of all players, unlike multi-player activity recognition, which recognizes separate actions in parallel. Its typical pipeline combines player tracking, individual features, and group-level information, and surveyed approaches use recurrent, graph, attention, and pose-based models.

IV. GROUP/TEAM ACTIVITY RECOGNITION

  • GAR recognizes a collective activity based on every player’s action and interactions, whereas multi-player activity recognition labels separate actions performed in parallel.The paper notes that evidence from all subjects may be needed to infer a group event such as a double-block.Source: §IV · p. 13
  • The typical GAR framework combines player tracking, individual-player feature extraction, and group feature combination.The survey describes this as more complicated than individual action recognition.Source: Figure 6 · p. 14
  • 51.1% on HierVolleyball and 81.9% on HierVolleyball-v2 were reported for Ibrahim et al.’s hierarchical LSTM; CERN reached 83.6% on HierVolleyball-v2.Ibrahim et al. aggregate player dynamics with a group-level LSTM, while CERN models player trajectories and relations with node and edge LSTMs.Source: §IV · p. 14
  • 87.1% accuracy on HierVolleyball-v2 was reported for Bagautdinov et al.’s end-to-end model, which shares a fully connected CNN across detection and recognition.RCRG extends two-stage approaches with a hierarchical relational network that aggregates information from neighboring graph nodes.Source: §IV · p. 14
  • StageNet combines player detection, semantic graph construction, temporal integration, and spatial-temporal attention.Its graph represents players as nodes and their spatial-distance and temporal-correlation relationships as edges; the attention module is described as making the model more explainable.Source: §IV · p. 14
  • Table VI compares deep-learning GAR models, mostly on HierVolleyball-v2, and the surveyed models are described as transferable to football and basketball.A football team-activity study used synchronized multi-camera video.Source: Table VI · p. 14, §IV · pp. 14–15
  • POGARS achieved 93.2% accuracy on HierVolleyball-v2, rising to 93.9% with player poses and ball tracklets; Pose3D achieved 91.3% accuracy.POGARS represents each player with 16 2D keypoints, while Pose3D uses skeleton heatmaps and a 3D CNN.Source: §IV · p. 14
  • 93.1% on HierVolleyball-v2 was reported for DIN, while Anchor-Transformer achieved 94.4% using pose and optical flow.DIN models player relations and dynamic walks in a spatial-temporal graph; Anchor-Transformer fuses player representations with a transformer.Source: §IV · p. 14

V. APPLICATIONS

The survey describes sports action recognition as a basis for training aids, video judging, highlight creation, sports news generation, and research. These applications use recognized actions to support analysis, assessment, and video content production.

Training Aids

  • Recognized action sequences can be related to winning strategies to guide player training and game-plan design.The survey also describes action recognition as a key step in AI coaching systems and a source of information for assessing player performance.Source: §V · p. 15

Game Assistance (Video Judge)

  • Sports video judges use action recognition as an elementary module for evaluating performances or recognizing whether referee trainees make proper signals.Examples include diving assessment and hand-gesture recognition for referee training.Source: §V · p. 15

Video Highlights

  • Recognized poses and actions can locate and stitch target moments into sports highlights.The survey describes figure-skating highlight detection and approaches that consider action diversity when constructing highlight videos.Source: §V · p. 15

Automatic Sports News Generation (ASNG)

  • Action recognition can automatically generate match statistics for sports news, reducing reliance on people to provide the numbers.Combined with visual captioning, recognition is also needed for natural, fluent, and accurate textual descriptions of video.Source: §V · p. 15

General Research Purposes

  • Sports action recognition supports research on athlete performance, rehabilitation, multimedia digests, and sports statistics analysis.The survey notes that sports datasets are used for benchmarking and that motion-capture systems support fine-grained gesture and pose analysis.Source: §V · p. 15

VI. CHALLENGES

The survey identifies practical challenges in sports action recognition, including costly data collection and annotation, fast and dense actions, changing camera views and occlusion, imbalanced classes, and multi-view video. It concludes by reviewing methods and open issues across sports and releasing a public-research toolbox.

VI. CHALLENGES

  • Sports datasets are harder to establish because broadcast footage can be inaccessible, self-recorded footage may be low quality, and fine-grained labels require domain expertise.Datasets serving multiple tasks require varied labels and auxiliary features; extracting table-tennis skeletons is difficult because strokes are dense and fast-moving.Source: §VI · p. 16
  • Table-tennis strokes commonly last 0.4–2 seconds, and 8–10 strokes can occur in under 6 seconds, challenging recognition of short events and action boundaries.The survey reports that performance on dense, fast-moving actions remains below expectations even after careful fine-tuning.Source: §VI · p. 16
  • Camera motion, cuts, and occlusion make motion features inconsistent, while existing occlusion-handling approaches are limited to single targets in multi-player sports.The surveyed benchmarks and traditional motion-trajectory methods commonly assume fixed cameras and clear views.Source: §VI · p. 16
  • Long-tailed class distributions and imbalanced data in soccer, basketball, and table tennis degrade model performance.The survey points to data augmentation or more robust recognition algorithms as possible ways to mitigate the problem; Figure 7 illustrates class distributions in P2A and SVW.Source: §VI · pp. 16–17, Figure 7 · p. 17
  • Most action-recognition benchmarks focus on single-camera or single-view actions, unlike sports videos, where multi-camera and multi-view robustness is needed.Training separate models for each view and combining results does not resolve the challenge of detecting and switching models across a complete video.Source: §VI · p. 17
  • Expensive fine-grained data collection motivates pretraining across sports and transferring to target tasks with few labeled samples or domain knowledge.The survey identifies few-shot and zero-shot learning as approaches requested for generalizing sports action recognition tasks.Source: §VI · p. 17

VII. CONCLUSION

  • The survey reviews methods and performance across team and individual sports, discusses open issues, and releases a toolbox for public sports-video research.Its coverage includes traditional statistical methods, 2D and 3D deep models, and skeleton-based methods using auxiliary information.Source: §VII · p. 17

Related papers

Loading 2206.01038v1…