MultiSports / README.md
JudieTest's picture
UPDATE: finished first version
477bcaa
metadata
annotations_creators:
  - crowdsourced
language:
  - en
language_creators:
  - expert-generated
license:
  - cc-by-nc-4.0
multilinguality:
  - monolingual
pretty_name: MultiSports
size_categories: []
source_datasets:
  - original
tags:
  - video
  - action detection
  - spatial-temporal action localization
task_categories:
  - image-classification
  - object-detection
  - other
task_ids:
  - multi-class-image-classification
extra_gated_heading: Acknowledge license to accept the repository
extra_gated_prompt: >-
  This work is licensed under a Creative Commons Attribution-NonCommercial 4.0
  International License
extra_gated_fields:
  I agree to use this dataset for non-commerical use ONLY: checkbox

Dataset Card for MultiSports

Table of Contents

Dataset Description

Dataset Summary

Spatio-temporal action localization is an important and challenging problem in video understanding. Previous action detection benchmarks are limited in aspects of small numbers of instances in a trimmed video or low-level atomic actions. MultiSports is a multi-person dataset of spatio-temporal localized sports actions. Please refer to this paper for more details. Please refer to this repository for evaluation.

Supported Tasks and Leaderboards

  • Spatial-temporal action localization

Details about evaluation can be found in the GitHub Repository. Previous challenge results can be found in this page and this CodaLab challenge.

Languages

The class labels in the dataset are in English.

Dataset Structure

Data Instances

Demo is available on dataset homepage.

The dataset contains rawframes.tar and multisports_GT.pkl. The GT pkl file is a dictionary with the following structure:

{
    'labels': ['label1', 'label2', ...],
    'train_videos': [['train_vid_1', 'train_vid_2', ...]],
    'test_videos': [['test_vid_1', 'test_vid_2', ...]],
    'nframes': {
        'vid_1': nframes_1,
        'vid_2': nframes_2,
        ...
    },
    'resolution': {
        'vid_1': resolution_1,
        'vid_2': resolution_2,
        ...
    },
    'gttubes': {
        'vid_1': {
            'label_1': [tube_1, tube_2, ...],
            'label_2': [tube_1, tube_2, ...],
            ...
        }
        ...
    }
}

Here a tube is a numpy.ndarray with nframes rows and 5 columns <frame number> <x1> <y1> <x2> <y2>.

Data Fields

Raw frames are organized according to their sport category. The pickle file of GT contains the following fields.

  • labels: list of labels

  • train_videos: a list with one split element containing the list of training videos

  • test_videos: a list with one split element containing the list of validation videos

  • nframes: dictionary that gives the number of frames for each video

  • resolution: dictionary that output a tuple (h,w) of the resolution for each video

  • gttubes: dictionary that contains the gt tubes for each video. Gt tubes are dictionary that associates from each index of label, a list of tubes. A tube is a numpy.ndarray with nframes rows and 5 columns <frame number> <x1> <y1> <x2> <y2>.

Please note that the label index starts from 0 and the frame index starts from 1. For the label index i, the label name is labels[i].

Click here to see the full list of MultiSports class labels mapping:
id Class
0 aerobic push up
1 aerobic explosive push up
2 aerobic explosive support
3 aerobic leg circle
4 aerobic helicopter
5 aerobic support
6 aerobic v support
7 aerobic horizontal support
8 aerobic straight jump
9 aerobic illusion
10 aerobic bent leg(s) jump
11 aerobic pike jump
12 aerobic straddle jump
13 aerobic split jump
14 aerobic scissors leap
15 aerobic kick jump
16 aerobic off axis jump
17 aerobic butterfly jump
18 aerobic split
19 aerobic turn
20 aerobic balance turn
21 volleyball serve
22 volleyball block
23 volleyball first pass
24 volleyball defend
25 volleyball protect
26 volleyball second pass
27 volleyball adjust
28 volleyball save
29 volleyball second attack
30 volleyball spike
31 volleyball dink
32 volleyball no offensive attack
33 football shoot
34 football long pass
35 football short pass
36 football through pass
37 football cross
38 football dribble
39 football trap
40 football throw
41 football diving
42 football tackle
43 football steal
44 football clearance
45 football block
46 football press
47 football aerial duels
48 basketball pass
49 basketball drive
50 basketball dribble
51 basketball 3-point shot
52 basketball 2-point shot
53 basketball free throw
54 basketball block
55 basketball offensive rebound
56 basketball defensive rebound
57 basketball pass steal
58 basketball dribble steal
59 basketball interfere shot
60 basketball pick-and-roll defensive
61 basketball sag
62 basketball screen
63 basketball pass-inbound
64 basketball save
65 basketball jump ball

Data Splits

train validation test
# of tubes 28514 10116 -

GT for test split is not provided. Please wait for the new competition to start. Information will be updated in dataset homepage.

Dataset Creation

Curation Rationale

Spatio-temporal action detection is an important and challenging problem in video understanding. Previous action detection benchmarks are limited in aspects of small numbers of instances in a trimmed video or low-level atomic actions.

Source Data

Initial Data Collection and Normalization

After choosing the four sports, we search for their competition videos by querying the name of sports like volleyball and the name of competition levels like Olympics and World Cup on YouTube, and then down- load videos from top search results. For each video, we only select high-resolution, e.g. 720P or 1080P, competition records and then manually cut them into clips of minutes, with less shot changes in each clip and to be more suitable for action detection.

Who are the source language producers?

The annotators of action categories and temporal boundaries are professional athletes of the corresponding sports. Please refer to the paper for more information.

Annotations

Annotation process

  1. (FIRST STAGE) A team of professional athletes generate records of the action la- bel, the starting and ending frame, and the person box in the starting frame, which can ensure the efficiency, accu- racy and consistency of our annotation results.

  2. At least one annotator with domain knowledge double-check the annotations, correct wrong or inaccurate ones and also add missing annotations

  3. (SECOND STAGE) With the help of FCOT tracking algorithm, a team of crowd-sourced annotators adjust bounding boxes of tracking results at each frame for each record.

  4. Double-check each instance by playing it in 5fps and manually correct the inaccurate bounding boxes.

Who are the annotators?

For the first stage, annotators are professional athletes. For the second stage, annotators are common volunteers.

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

Authors of this paper

  • Yixuan Li

  • Lei Chen

  • Runyu He

  • Zhenzhi Wang

  • Gangshan Wu

  • Limin Wang

Licensing Information

Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.

Citation Information

If you find this dataset useful, please cite as

@InProceedings{Li_2021_ICCV,
    author    = {Li, Yixuan and Chen, Lei and He, Runyu and Wang, Zhenzhi and Wu, Gangshan and Wang, Limin},
    title     = {MultiSports: A Multi-Person Video Dataset of Spatio-Temporally Localized Sports Actions},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {13536-13545}
}

Contributions

Thanks to @Judie1999 for adding this dataset.