tnguyen2002/overwatch2-killfeed-rows

Overwatch 2 Killfeed Rows

Bounding-box annotations of killfeed rows in Overwatch 2 gameplay footage.

A killfeed row is the banner that appears top-right when a player is eliminated, showing the attacker, the ability used, and the victim. This dataset marks where those rows are — one class, one box per row.

It exists because reading the killfeed is the most reliable cheap way to find kills in a VOD. Frame-differencing and audio cues both fire on the wrong things; the killfeed is unambiguous and always in the same region of the screen.

Contents

Split	Images	Boxes	Empty	Boxes/image
train	385	535	78	1.39
validation	23	30	8	1.30
test	45	61	17	1.36
total	453	626	103	1.38

One class: kill_row. About 10 MB.

103 images contain no rows at all. Those are deliberate negatives — most of a match has an empty killfeed, and a detector trained only on positives learns to hallucinate rows in quiet frames.

Two resolutions, from different source recordings: 730×453 (185 images) and 487×302 (268).

Format

HuggingFace imagefolder layout — one directory per split, each with its images and a metadata.jsonl.

{
  "file_name": "frame_0001_png.rf.<hash>.jpg",
  "width": 487, "height": 302,
  "objects": {"bbox": [[296.0, 22.0, 171.0, 28.0]], "category": [0]}
}

Boxes are COCO-style [x, y, width, height] in pixels, origin top-left.

from datasets import load_dataset
ds = load_dataset("<repo-id>")
ds["train"][0]["objects"]     # {'bbox': [[...]], 'category': [0]}

How it was collected

1. Frames sampled from Overwatch 2 VODs at 0.5 fps (one per 2s). A row stays on screen 5–8 seconds, so every kill is caught several times.
2. Each frame cropped to the killfeed band — the top-right region at (x0, x1, y0, y1) = (0.62, 1.00, 0.00, 0.42) in frame fractions. Deliberately wider and taller than the feed so stacked rows and HUD-scale variation stay inside it.
3. Rows annotated by hand as a single kill_row class.

Cropping to the band rather than labelling full frames is what makes a set this size workable: the model never has to learn to ignore 60% of the screen.

Contents note: gamertags are visible

Killfeed rows show players' in-game display names, and they have not been redacted. They are pseudonymous handles, visible to everyone in the match and in any publicly posted clip — but they are present in these images, and a dataset is mirrored and machine-readable in a way a screenshot is not. Flagged so anyone using it can make their own call.

The kill_row task does not use them: the labels locate rows, they do not read them. The names are incidental to the annotation, not part of it.

Intended use

Training a detector to locate killfeed rows — the first stage of a pipeline that then OCRs the attacker name and attributes the kill to a specific player.

Not suitable for reading names or hero identity (no labels for either), broadcast/observer footage, or games other than Overwatch 2.

Limitations

• Single class. A row is a row. No label for who got the kill, which hero was used, or whether it was a final blow or an assist.
• POV-biased. Almost entirely first-person player footage. Broadcast and observer overlays render the feed differently and are under-represented.
• Small. 453 images is enough to fine-tune a detector on a fixed, high-contrast UI element; not enough to generalise across game patches.
• Two HUD scales only. Overwatch lets players resize the HUD, and this set spans two source recordings, not that whole range.
• Patch-bound. Blizzard has restyled the killfeed before and will again. These labels describe the UI as it looked when the footage was recorded.
• One annotation was traced as a polygon rather than a rectangle; it was converted to its bounding extent during export.

Licence and attribution

Annotations are released under CC BY 4.0, matching the source Roboflow project.

The images contain Overwatch 2 UI and hero art, which is © Blizzard Entertainment. This dataset is published for research use and is not endorsed by or affiliated with Blizzard; the CC BY licence covers the annotations, not the underlying game imagery.

Citation

@misc{overwatch2_killfeed_rows,
  title  = {Overwatch 2 Killfeed Rows},
  author = {Tom Nguyen},
  year   = {2026},
  url    = {https://huggingface.co/datasets/tnguyen2002/overwatch2-killfeed-rows}
}