BlackPymer2/BrawlStars-gameplay

Brawl Stars Gameplay Dataset & Models

This repository contains labeled gameplay footage of Brawl Stars, along with trained computer vision models for detection and HP (health points) OCR.

Contents

YOLO Detection Model

• —`best.pt` — Trained YOLO model weights (yolo26s architecture). Detects: player, enemy, powercube, powercube-box, wall, water, bush, zone.

from ultralytics import YOLO
model = YOLO("best.pt")
results = model.predict("screenshot.png")

HP OCR Dataset

health_dataset/ — A labeled dataset of 3259 cropped HP regions extracted from gameplay frames:

• —images/ — 100×115 px cropped HP bars (3-channel)
• —labels.csv — Ground truth HP values (regression targets)
• —labels_with_predictions.csv — Labels augmented with model predictions for unlabelled entries

Each filename encodes metadata: {source}_frame{number}_{player|enemy}_{instance_id}.jpg

Distribution

HP OCR Model

• —`hp_crnn_best.pt` — CRNN (CNN + BiLSTM + CTC) for reading HP values from cropped health bars.
• —Accuracy: 93.75% on the validation split.

[image]

Loading & Inference

import cv2
import torch
from train_hp import CRNN, preprocess, decode
model = CRNN()
model.load_state_dict(torch.load("hp_crnn_best.pt", map_location="cpu"))
model.eval()
# Load a cropped HP region
img = cv2.imread("health_dataset/images/frame_000306_player_0.jpg")
# Preprocess (grayscale, resize to H=48, normalize)
inp = torch.from_numpy(preprocess(img)).unsqueeze(0)
# Inference
with torch.no_grad():
    logits = model(inp).permute(1, 0, 2)
pred_ids = logits.argmax(dim=2).permute(1, 0)[0]
hp_text = decode(pred_ids.tolist())
print(f"HP: {hp_text}")

Training

The model is a lightweight CRNN trained via CTC loss on the HP dataset. See train_hp.py for the full training pipeline.

Video Demo

A video demonstrating the trained network in action is included in this repository.

License

Apache 2.0