overwatch2-killfeed-detector

Overwatch 2 Killfeed Row Detector

A YOLOv8s detector that finds killfeed rows in Overwatch 2 gameplay — the banners that appear top-right when a player is eliminated.

Locating the row is the first stage of finding kills in a VOD. Once you have the box you can OCR the attacker's name and attribute the kill; without it you are guessing from frame differences or audio, both of which fire on the wrong things.

Results

Evaluated on the dataset's held-out test split (45 images, 61 boxes) — images never seen in training or validation.

Inference measured at ~25 ms/image at 640px on Apple MPS; substantially faster on a CUDA GPU.

The high mAP@50 is expected and not as impressive as it looks: a killfeed row is a fixed, high-contrast UI element in a predictable screen region, which is close to the easiest case object detection has. The gap down to 0.83 at mAP@50-95 is the honest number — box edges are loose, because a row's exact boundary is ambiguous where the coloured bar fades into its border.

Usage

from ultralytics import YOLO

model = YOLO("killfeed.pt")
results = model.predict("frame.png", imgsz=640, conf=0.25)

for box in results[0].boxes:
    x1, y1, x2, y2 = box.xyxy[0].tolist()
    print(f"kill_row at ({x1:.0f}, {y1:.0f}) - ({x2:.0f}, {y2:.0f}) conf={box.conf.item():.2f}")

An ONNX export (killfeed.onnx) is included for runtimes without PyTorch.

Feed it the killfeed band, not the whole frame. The model was trained on crops of the top-right region — (x0, x1, y0, y1) = (0.62, 1.00, 0.00, 0.42) as fractions of the frame. Passing a full 1920×1080 screenshot will work poorly, because at 640px inference the row shrinks to a handful of pixels.

Training

The released checkpoint has had its optimizer and EMA state stripped (89.5 MB → 22.5 MB); it scores identically to the full checkpoint.

Classes

One: kill_row.

The model finds rows. It does not identify who got the kill, which hero was used, or whether a row is a final blow or an assist — those are downstream problems and there are no labels for them here.

Limitations

• POV footage. Trained almost entirely on first-person player recordings. Broadcast and observer overlays render the killfeed differently.
• Two HUD scales. Overwatch lets players resize the HUD; the training data spans two source recordings, not that whole range.
• Patch-bound. Blizzard has restyled the killfeed before and will again. This model describes the UI as it looked in mid-2026.
• Band-cropped input. See usage above — this is a real constraint, not a preference.
• Small training set. 453 images. Enough for a fixed UI element, not enough to survive a visual redesign.

Licence

AGPL-3.0, inherited from Ultralytics YOLOv8, which this model was trained with and which its weights derive from. Ultralytics offers a separate Enterprise licence for use without AGPL obligations.

Note the network clause: AGPL-3.0 requires that users interacting with the software over a network be able to obtain its corresponding source. If you deploy this model in a service, that applies to you.

The training images contain Overwatch 2 UI and hero art, © Blizzard Entertainment. This model is published for research use and is not endorsed by or affiliated with Blizzard.

Citation

@misc{overwatch2_killfeed_detector,
  title  = {Overwatch 2 Killfeed Row Detector},
  author = {Tom Nguyen},
  year   = {2026},
  url    = {https://huggingface.co/tnguyen2002/overwatch2-killfeed-detector}
}