🏐 RIT-18 Dataset

Contents

0. Overview
1. Download
2. Details of Dataset
3. Benchmark
4. Citation

🌐 Overview

RIT-18 is a novel compositional activity dataset collected by ACTION Lab at RIT containing 18 compositional activity classes. We collected video clips from 51 volleyball games on YouTube. With comprehensive annotations, RIT-18 is a large scale dataset for group activity understanding tasks such as group activity recognition, future activity anticipation, and rally-level winner prediction. The benchmark for these three tasks is provided.

Besides the aforementioned tasks, we believe RIT-18 dataset is potential for other less explored tasks, for example, temporal group activity localization, individual contribution evalution, and game-level winner prediction. Welcome to play!

🗄️ Download

Download here

🗂️ Details of Dataset

Overall structure of the dataset

RIT-18
├── 0                  # folder for video 0
├── 1
│   ├── clips          # folder for clips
│   │   ├── 000.mp4
│   │   ├── ...
│   │   └── 029.mp4
│   ├── group.txt      # activity label for all the clips of the video
│   └── info.txt       # match description, video copyright, resolution and other format information
├── ...
├── 50
├── copyright.txt      # all the links to the videos
└── annotation.pkl     # all the annotations

Activity Labels

We annotated 12035 frames picked from 1530 clips (51 videos * 30 clips / video) with 18 action labels.

Activity Labels	No. of Instances	Activity Labels	No. of Instances
l-serve	774	r-serve	756
l-firstpass	1395	r-firstpass	1395
l-set	1270	r-set	1266
l-spike	1012	r-spike	1014
l-volley	136	r-volley	143
l-drop	243	r-drop	235
l-shot	23	r-shot	23
l-block	393	r-block	427
l-winpoint	764	r-winpoint	766

Annotation Format

group.txt

[clip-name].mp4
[a b] [a b] [left/right]
[key-frame-num] [start-frame-num] [end-frame-num] [key-time] [start-time] [end-time] [activity-label] [bounding box]
[key-frame-num] [start-frame-num] [end-frame-num] [key-time] [start-time] [end-time] [activity-label] [bounding box]
[key-frame-num] [start-frame-num] [end-frame-num] [key-time] [start-time] [end-time] [activity-label] [bounding box]
[key-frame-num] [start-frame-num] [end-frame-num] [key-time] [start-time] [end-time] [left/right winpoint]
    
e.g.
    000.mp4
    0 0 0 1 left
    000000 000000 000026 0.000 0.000 0.520 r-serving 1073 330 1147 533
    000052 000027 000101 1.040 0.540 2.010 l-passing 355 382 422 589
    000124 000102 000164 2.480 2.030 3.285 l-setting 513 426 574 623
    000182 000165 000201 3.640 3.305 4.020 l-spiking 472 381 541 601
    000202 000202 000247 4.040 4.040 4.940 l-winpoint

info.txt

startTeam: [left/right]
winner: [left/right]
score: [a b]
1: [a b]
2: [a b]
3: [a b]
4: [a b]
link: [download link]
frameRate=[fps]
imWidth=[frame image width]
imHeight=[frame image height]
imExt=[frame image extension]

# 'left' and 'right' stand for the team position when the match or video begins
# 'a' is the score of 'left'
# 'b' is the score of 'right'

e.g. 
    startTeam: right
    winner: right
    score: 3 0
    1: 25 22
    2: 28 26
    3: 26 24
    link: https://www.youtube.com/watch?v=A1B2C3D4E
    frameRate=25
    imWidth=1920
    imHeight=1080
    imExt=.jpg

🚧 Benchmark

Group activity recognition

Group acitivity anticipation

Rally-level winner prediction

Winner Prediction baseline test.

cd winner
python scripts/test_volleyball_winner.py

Our Poster

📑 Citation

Please cite our CVPR Workshop paper if use this dataset.

@InProceedings{RIT18_2020_CVPR_Workshops,
    author = {Chen, Junwen and Hao, Haiting and Hong, Hanbin and Kong, Yu},
    title  = {RIT-18: A Novel Dataset for Compositional Group Activity Understanding},
    booktitle = {IEEE Conference on Computer Vision and Pattern Recognition Workshop},
    month  = {June},
    year   = {2020}
}