The Brazilian Jiu-Jitsu Positions Dataset is a dataset for training jiu-jitsu position classification methods. The dataset contains 120.279 labeled images of 2 jiu-jitsu athletes sparring in different combat positions. The poses of the athletes were detected automatically. They were, however, manually verified and the best detected poses were selected. Nevertheless, the correctness of the poses is not guaranteed. The combat positions were labeled manually. The dataset includes images of 10 positions, resulting in 18 classes.

10 combat positions and 18 classes:

• Standing
• Takedown - 2 classes designating which athlete is initiating the take-down
• Open guard - 2 classes designating which athlete is currently in the guard position (We do not distinguish between the type of open guard for example both “de la RIva” and “spider” guards are considered open guard.)
• Half guard - 2 classes
• Closed guard - 2 classes
• 50-50 guard (an aggregate of all positions where both athletes are in the guard position, from leg entanglements, to double guard)
• Side control - 2 classes designating which athlete is in the top position (north-south and knee-on belly positions are treated as side control in this dataset)
• Mount - 2 classes designating which athlete is in the top position
• Back - 2 classes designating which athlete is controlling the opponents back (despite most rule-sets stating hooks - controlling the opponents rotation with your legs, are required to obtain the back position, this is not ensured in the dataset)
• Turtle - 2 classes designating which athlete is in the top position

The dataset was captured using 3 smart phone cameras, shooting 6 sparring sequences. We selected the images in which our pose tracking framework detected at least one athlete’s pose and a position label was available. The images are named with the following convention: the first 2 digits identify the video sequence, and the last 5 the position of the frame in the video. The poses of the athletes are in MS-COCO format, consisting of 17 keypoints representing the athlete’s joints (nose, left eye, right eye, left ear, right ear, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, right ankle). Each joint is represented by its x and y coordinates in the image and the confidence of the detector—[x, y, c]. All annotations are stored in a json array of objects.

Each annotation contains the following values:

• Image - the name of the image (5 digits - VSFRAME - Video Sequence, FRAME)
• Pose1 - pose of the athlete marked as 1 in the above format ([[x0, y0, c0]… [x16, y16, c16]])
• Pose2 - pose of the athlete marked as 2 in the above format
• Position - the code for the combat position of the athlete (eg. mount2 - athlete 2 is in mount top position)
• Frame - number of the frame in the video ( all videos have roughly 30 fps)

DOWNLOAD IMAGES

DOWNLOAD ANNOTATIONS

License and citation

The dataset is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Please cite our paper published in Video-Based Detection of Combat Positions and Automatic Scoring in Jiu-jitsu

@inproceedings{hudovernik2023MMW,
  author = {Hudovernik, Valter and  Sko{\v{c}}aj, Danijel},
  title = {{Video-Based Detection of Combat Positions and Automatic Scoring in Jiu-jitsu}},
  booktitle = {Proceedings of the ACM Multimedia Workshop MMSports’22},  
  year = {2022},
  month = {October},
  day = {10},
  location = {Lisboa, Portugal}
}