HSLVision-Depth

The depth companion to HSLVision, the first public object detection dataset and benchmark for the RoboCup Humanoid Soccer League (HSL). Every one of the 6,068 images is paired with a pixel-aligned metric depth map, turning the RGB release into a full RGB-D benchmark with 65,309 bounding-box annotations across 7 classes.

The depth maps share the same pixel coordinates as the RGB images, so the object detection annotations apply directly to both modalities without transformation.

🌐 Project page: https://intelligentroboticslab.github.io/hslvision/

🖼️ RGB-only dataset: whirlwind-ams/hslvision

Classes

Annotations cover seven object categories (category id 0 is reserved as background):

id	Name	Description
1	Ball	The soccer ball
2	Goalpost	Goalposts (designs vary by venue)
3	K1	Humanoid robots
4	L-Intersections	Field-line L-intersections
5	Penalty Marks	Penalty marks
6	T-Intersections	Field-line T-intersections
7	X-Intersections	Field-line X-intersections

Data collection

Images were collected across six locations spanning a range of field layouts, goalpost designs, ball sizes, robot platforms, and HSL field sizes (small, middle, large):

• LAB42: whIRLwind
• German Open 2026, Cologne: whIRLwind
• RCAP Beijing Masters 2025: whIRLwind, HTWK
• RoboCup 2025, Salvador: HTWK
• RoboCup 2019, Sydney: BitBots
• RoboCup 2017, Nagoya: BitBots

Frame selection

Raw footage is subsampled at 3 fps and filtered to remove motion blur, occlusion, and low-quality frames. To avoid near-identical, redundant images, each frame is embedded with a pretrained DINOv3 ViT-S/16 encoder (384-d L2-normalized pooler token). Images with the most above-threshold neighbors are iteratively removed until no pair exceeds a cosine-similarity threshold τ, yielding a diverse subset independent of frame order.

Figure 1. Overview of the data-selection pipeline. Embedding-space pruning iteratively removes points with many nearby neighbors; the effect of varying τ on the diversity of the selected subset is shown at the bottom.

Depth maps

Onboard stereo depth is not available across all recording platforms, so every color frame is paired with a metric depth map estimated by Depth Anything V3. The depth maps share the same pixel coordinates as the RGB images, so annotations apply directly to both modalities without transformation.

Figure 2. RGB input, Hobot StereoNet, and Depth Anything V3 predictions across two scenes. Depth Anything V3 produces smoother, more spatially consistent depth with sharp, well-localized object boundaries.

Usage

from datasets import load_dataset

ds = load_dataset("whirlwind-ams/hslvision-depth")
example = ds["train"][0]

image = example["image"]              # PIL.Image, RGB frame
depth = example["metric_depth"]       # PIL.Image, pixel-aligned metric depth
objects = example["objects"]
for bbox, category in zip(objects["bbox"], objects["category"]):
    x, y, w, h = bbox                 # COCO [x, y, width, height], valid for both modalities
    print(category, bbox)

Citation

@inproceedings{catarrinho2026hslvision,
  author    = {Xavier Catarrinho, D. M. and de Jong, G. and Meijer, M. J. and Ruiter, H. and Honkoop, M.},
  title     = {HSLVision: A Multimodal Vision Dataset for RoboCup Humanoid Soccer},
  booktitle = {RoboCup 2026: Robot World Cup XXIX},
  year      = {2026},
  location  = {Incheon, South Korea},
}

Acknowledgements

Created by team whIRLwind Amsterdam (University of Amsterdam).

Thanks to all whIRLwind members for their annotation efforts, and to team HTWK for providing some of the raw recordings used in this work.

License

Released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share and adapt the data, provided you give appropriate credit.

Hungry for more?

HSLVision is a whIRLwind project. Our mission is to push the boundaries of robotics and AI in robot football.

We are always looking for interested students or collaboration partners to join us. Learn more at whirlwind.team.