Sports is a small category in open data and a well-posed one: the rules define what counts as correct, which makes evaluation unusually clean.
What makes sports tractable is that the semantics are specified. A pass either reached its target or did not; a player was either offside or was not. That removes the labelling ambiguity that dogs most vision tasks, and it means automated evaluation against recorded outcomes is possible in a way it rarely is elsewhere.
Sports data comes in three layers, and most models turn one into the next. Video is the raw layer. Tracking data is the position of every player and the ball, sampled many times a second; detectors and trackers produce it from video. Event data is one row per action on the ball — a pass, a shot, a tackle — with its time, location and outcome; action-spotting models produce it, and almost all analytics is built on it. Decide which layer your model consumes and which it outputs before you look for data, because the datasets split along the same lines.
The constraint is rights. Broadcast footage is commercially licensed and professional tracking data is proprietary, so open sports data tends to be small, amateur, or derived: positions and events extracted from footage rather than the footage itself. Two well-known open resources show the pattern. StatsBomb's open data publishes event data for selected competitions for research and analysis, on condition that you credit StatsBomb. SoccerNet publishes its action-spotting annotations openly, but the broadcast videos they label are released only after you sign a non-disclosure agreement. Games are the exception: Lichess releases its entire game database under CC0, which is why chess is the cleanest data in this category.
The six sets below cover the layers that open data reaches: player detection in match footage, action recognition (tennis strokes, and UCF101's sports-heavy action classes), player identity, and game state for chess and cards. Before you build on any of the video or image sets, check where their frames came from — a dataset that was cut from broadcasts inherits the broadcaster's rights, whatever licence its uploader chose. For commercial work this is the category where you are most likely to need a data licence rather than a dataset, and that is a procurement question rather than a search problem.
| Dataset | Modality | Size | Licence | Best for |
|---|---|---|---|---|
| Tennis Player Actions Dataset | Image | 2,004 images | CC BY 4.0 (source unclear) | Recognition of sports actions |
| Football Players Detection | Image | 106 images | CC BY 4.0 (footage unclear) | Object detection to track players |
| UCF101 | Video | 13,320 videos | Not stated | Action recognition with 3D CNNs |
| Golden Foot Football Players Dataset | Image | 7,188 images | Unclear — see below | Facial recognition applied |
| Chess Checkmate Images | Image | 3,500 images | Apache 2.0 | Vision models understand chess dynamics |
| Playing Cards Classification Dataset | Image | 8,157 images | CC0 (source unclear) | Image classification |
Tennis Player Actions Dataset
- Size
- 2,004 images + JSON COCO keypoints, 4 classes, ~508 MB
- Licence
- CC BY 4.0 as uploaded by a re-uploader; the original source is not named and the images show identifiable people
Over 2,000 images of four tennis actions, forehand, backhand, serve and waiting position, about 508 MB, annotated in COCO JSON format with 18 OpenPose body keypoints per figure. Pose keypoints rather than boxes alone.
Training recognition of sports actions from still images, analysing body movement via fine keypoint annotation, and developing coaching or automatic scoring assistants.
Source: innovatiana.com
Football Players Detection
- Size
- 106 annotated JPG images in YOLO format (.txt), 375 files in total
- Licence
- CC BY 4.0 as uploaded; a dataset of the same name was built from Bundesliga broadcast clips, so the footage rights are not cleared
106 annotated JPG sports-scene images with YOLO-format TXT labels locating each soccer player on the field, 375 files in total. Extremely small, so really a demonstration or fine-tuning starter.
Training object detection to track players on the field, developing sports and live-stream video analysis, and learning object detection algorithms such as YOLO.
Source: innovatiana.com
UCF101
- Size
- 13,320 videos classified into 101 categories of human actions, AVI format
- Licence
- No licence stated on the official page; the clips come from YouTube (the catalogue card says CC BY-NC 4.0)
13,320 short AVI clips spanning 101 human activity classes, the long-standing baseline for action recognition. The official page states no licence and the clips are third-party YouTube videos, so the footage is not cleared for commercial use. Now small enough to serve mainly as a fast comparison point.
Training action recognition with 3D CNNs, RNNs or video transformers, validating embedded vision systems, and researching space-time architectures.
Source: innovatiana.com
Golden Foot Football Players Dataset
- Size
- 7,188 images in JPG format, divided by player
- Licence
- Labelled CC0 by the uploader; portraits of real players with no stated source
7,188 JPG portraits of footballers who have won the Golden Foot award since 2003, organised into one folder per player. Identity-labelled face data in a sports context. The CC0 tag is the uploader's, the photographs have no stated source, and they show identifiable people, so neither copyright nor privacy is settled.
Training facial recognition applied to sport, creating a visual search engine for player profiles, and testing detection and identification algorithms.
Source: innovatiana.com
Chess Checkmate Images
- Size
- 3,500 images in PNG format, approx. 83 MB
- Licence
- Apache-2.0
3,500 PNG images of chess board positions taken from real games, each captured at the moment just before a checkmate. A third-party visual version of a BIG-bench task rather than part of BIG-bench itself, framed as a reasoning probe rather than plain classification.
Testing whether vision models understand chess dynamics, training multimodal models combining vision and logical reasoning, and creating agents that analyse games from images.
Source: innovatiana.com
Playing Cards Classification Dataset
- Size
- 8,157 JPEG images (224×224), 53 classes
- Licence
- CC0 as uploaded; the image source is not stated
8,157 clean, well-centred JPEG images at 224x224 across 53 classes, one per standard playing card plus the joker. Uniform framing and sizing make it an easy, reliable fine-grained classification exercise.
Training image classification such as CNNs or MobileNet, creating card recognition for automated games, and beginner computer vision teaching projects.
Source: innovatiana.com
How this list was put together
Drawn from the datasets in Dataset Finder filed under this topic, then checked by hand against what this page covers. A tag is a filing decision, not a description, so records that carried the tag without being about the topic were removed, and relevant records filed elsewhere were added. The order puts the closest fit first.
Sizes and licences come from each dataset's own documentation, vague wording included. Where we found a re-upload or catalogue card stating a licence that the original release contradicts — a CC0 tag on Flickr photographs, MIT on data whose terms require a signed agreement — the entry gives the original terms and says so. The comparison table abbreviates licences; hover a cell for the full wording.