Ysobel/robot_jersey_dataset

NUpbr RoboCup Jersey-Colour Robot Detection Dataset

1,707 physically-based-rendered images of RoboCup Humanoid Soccer scenes, generated with NUpbr, each containing 1-4 robots and one ball, with per-instance bounding boxes and, for each robot, a jersey colour.

Every frame is restricted to at most two distinct jersey colours across all robots present, regardless of robot count. Each robot is independently assigned one of the two colours sampled fresh for that frame from a wide, vividly-saturated colour space. This makes the dataset suited to colour-conditioned role classification (e.g. teammate/opponent assignment from an externally supplied reference colour), in addition to standard robot/ball detection.

Robots whose bounding box is cropped at the top edge of the frame are excluded from the annotations: a top-cropped box typically shows only legs, with the jersey-bearing torso cut off entirely, leaving no colour signal to learn from. Images left with no remaining robots after this filter are dropped outright. 2,000 raw renders were filtered down to the 1,707 images and 2,702 robot instances published here.

Dataset structure

images/000001.png ... images/001707.png    1280x1024 RGBA renders
metadata.jsonl                              one JSON object per image

Each line of metadata.jsonl:

{
  "file_name": "images/000001.png",
  "robots": [
    {"bbox": [640, 70, 717, 240], "jersey_colour": "#c339e6", "model": "GankenKun"}
  ],
  "ball": {"bbox": [408, 470, 523, 585], "id": "ball_030"}
}

• bbox: [x_min, y_min, x_max, y_max] in integer pixel coordinates, origin top-left.
• jersey_colour: #rrggbb hex string, unique per robot-colour-group within a frame (at most two distinct values per image).
• model: the robot mesh/model used for that instance (e.g. nugus,GankenKun, darwin, wolfgang).
• ball.id: the ball texture/asset identifier used for that instance.
• Every image has exactly one ball and between one and four robots.

This follows the Hugging Face imagefolder convention -- loadable directly with:

from datasets import load_dataset
ds = load_dataset("imagefolder", data_dir="path/to/robot_jersey_dataset")

Generation

Images were rendered with NUpbr, a Blender-based physically-based rendering pipeline for RoboCup Humanoid Soccer scenes. Robot positions, camera viewpoint, HDRI environment, and jersey colours are randomised per frame; a frame is discarded and regenerated if it does not end up with between one and four visible robots, or if any visible robot or ball would otherwise be left without a valid bounding box (occluded, out of frame, or an implausible fisheye-projection artifact).

After rendering, robots whose box is cropped at the top edge of the frame are removed from the annotations (jersey not visible), and any image left with no annotated robots as a result is dropped from the release entirely.